Upload folder using huggingface_hub

src/__init__.py

@@ -0,0 +1,5 @@
+"""
+valtec-tts source package
+"""
+
+__version__ = "1.0.0"

src/alignment/__init__.py

@@ -0,0 +1,64 @@
+"""
+Monotonic alignment package
+"""
+
+import numba
+from numpy import zeros, int32, float32
+from torch import from_numpy
+
+
+@numba.jit(
+    numba.void(
+        numba.int32[:, :, ::1],
+        numba.float32[:, :, ::1],
+        numba.int32[::1],
+        numba.int32[::1],
+    ),
+    nopython=True,
+    nogil=True,
+)
+def maximum_path_jit(paths, values, t_ys, t_xs):
+    b = paths.shape[0]
+    max_neg_val = -1e9
+    for i in range(int(b)):
+        path = paths[i]
+        value = values[i]
+        t_y = t_ys[i]
+        t_x = t_xs[i]
+
+        v_prev = v_cur = 0.0
+        index = t_x - 1
+
+        for y in range(t_y):
+            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+                if x == y:
+                    v_cur = max_neg_val
+                else:
+                    v_cur = value[y - 1, x]
+                if x == 0:
+                    if y == 0:
+                        v_prev = 0.0
+                    else:
+                        v_prev = max_neg_val
+                else:
+                    v_prev = value[y - 1, x - 1]
+                value[y, x] += max(v_prev, v_cur)
+
+        for y in range(t_y - 1, -1, -1):
+            path[y, index] = 1
+            if index != 0 and (
+                index == y or value[y - 1, index] < value[y - 1, index - 1]
+            ):
+                index = index - 1
+
+
+def maximum_path(neg_cent, mask):
+    device = neg_cent.device
+    dtype = neg_cent.dtype
+    neg_cent = neg_cent.data.cpu().numpy().astype(float32)
+    path = zeros(neg_cent.shape, dtype=int32)
+
+    t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
+    t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
+    maximum_path_jit(path, neg_cent, t_t_max, t_s_max)
+    return from_numpy(path).to(device=device, dtype=dtype)

src/alignment/monotonic_align.py

@@ -0,0 +1,46 @@
+import numba
+
+
+@numba.jit(
+    numba.void(
+        numba.int32[:, :, ::1],
+        numba.float32[:, :, ::1],
+        numba.int32[::1],
+        numba.int32[::1],
+    ),
+    nopython=True,
+    nogil=True,
+)
+def maximum_path_jit(paths, values, t_ys, t_xs):
+    b = paths.shape[0]
+    max_neg_val = -1e9
+    for i in range(int(b)):
+        path = paths[i]
+        value = values[i]
+        t_y = t_ys[i]
+        t_x = t_xs[i]
+
+        v_prev = v_cur = 0.0
+        index = t_x - 1
+
+        for y in range(t_y):
+            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+                if x == y:
+                    v_cur = max_neg_val
+                else:
+                    v_cur = value[y - 1, x]
+                if x == 0:
+                    if y == 0:
+                        v_prev = 0.0
+                    else:
+                        v_prev = max_neg_val
+                else:
+                    v_prev = value[y - 1, x - 1]
+                value[y, x] += max(v_prev, v_cur)
+
+        for y in range(t_y - 1, -1, -1):
+            path[y, index] = 1
+            if index != 0 and (
+                index == y or value[y - 1, index] < value[y - 1, index - 1]
+            ):
+                index = index - 1

src/models/__init__.py

@@ -0,0 +1,5 @@
+"""
+TTS Models package
+"""
+
+from .synthesizer import SynthesizerTrn, Generator, MultiPeriodDiscriminator

src/models/synthesizer.py

@@ -0,0 +1,1030 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from src.nn import commons
+from src.nn import modules
+from src.nn import attentions
+
+from torch.nn import Conv1d, ConvTranspose1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+
+from src.nn.commons import init_weights, get_padding
+from src import alignment as monotonic_align
+
+
+class DurationDiscriminator(nn.Module):  # vits2
+    def __init__(
+        self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = modules.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = modules.LayerNorm(filter_channels)
+        self.dur_proj = nn.Conv1d(1, filter_channels, 1)
+
+        self.pre_out_conv_1 = nn.Conv1d(
+            2 * filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.pre_out_norm_1 = modules.LayerNorm(filter_channels)
+        self.pre_out_conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.pre_out_norm_2 = modules.LayerNorm(filter_channels)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+        self.output_layer = nn.Sequential(nn.Linear(filter_channels, 1), nn.Sigmoid())
+
+    def forward_probability(self, x, x_mask, dur, g=None):
+        dur = self.dur_proj(dur)
+        x = torch.cat([x, dur], dim=1)
+        x = self.pre_out_conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.pre_out_norm_1(x)
+        x = self.drop(x)
+        x = self.pre_out_conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.pre_out_norm_2(x)
+        x = self.drop(x)
+        x = x * x_mask
+        x = x.transpose(1, 2)
+        output_prob = self.output_layer(x)
+        return output_prob
+
+    def forward(self, x, x_mask, dur_r, dur_hat, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+
+        output_probs = []
+        for dur in [dur_r, dur_hat]:
+            output_prob = self.forward_probability(x, x_mask, dur, g)
+            output_probs.append(output_prob)
+
+        return output_probs
+
+
+class TransformerCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+        share_parameter=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+
+        self.wn = (
+            attentions.FFT(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=self.gin_channels,
+            )
+            if share_parameter
+            else None
+        )
+
+        for i in range(n_flows):
+            self.flows.append(
+                modules.TransformerCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    n_layers,
+                    n_heads,
+                    p_dropout,
+                    filter_channels,
+                    mean_only=True,
+                    wn_sharing_parameter=self.wn,
+                    gin_channels=self.gin_channels,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class StochasticDurationPredictor(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        filter_channels = in_channels  # it needs to be removed from future version.
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.log_flow = modules.Log()
+        self.flows = nn.ModuleList()
+        self.flows.append(modules.ElementwiseAffine(2))
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.flows.append(modules.Flip())
+
+        self.post_pre = nn.Conv1d(1, filter_channels, 1)
+        self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.post_convs = modules.DDSConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        self.post_flows = nn.ModuleList()
+        self.post_flows.append(modules.ElementwiseAffine(2))
+        for i in range(4):
+            self.post_flows.append(
+                modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.post_flows.append(modules.Flip())
+
+        self.pre = nn.Conv1d(in_channels, filter_channels, 1)
+        self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.convs = modules.DDSConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
+
+    def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
+        x = torch.detach(x)
+        x = self.pre(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.convs(x, x_mask)
+        x = self.proj(x) * x_mask
+
+        if not reverse:
+            flows = self.flows
+            assert w is not None
+
+            logdet_tot_q = 0
+            h_w = self.post_pre(w)
+            h_w = self.post_convs(h_w, x_mask)
+            h_w = self.post_proj(h_w) * x_mask
+            e_q = (
+                torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype)
+                * x_mask
+            )
+            z_q = e_q
+            for flow in self.post_flows:
+                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
+                logdet_tot_q += logdet_q
+            z_u, z1 = torch.split(z_q, [1, 1], 1)
+            u = torch.sigmoid(z_u) * x_mask
+            z0 = (w - u) * x_mask
+            logdet_tot_q += torch.sum(
+                (F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2]
+            )
+            logq = (
+                torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2])
+                - logdet_tot_q
+            )
+
+            logdet_tot = 0
+            z0, logdet = self.log_flow(z0, x_mask)
+            logdet_tot += logdet
+            z = torch.cat([z0, z1], 1)
+            for flow in flows:
+                z, logdet = flow(z, x_mask, g=x, reverse=reverse)
+                logdet_tot = logdet_tot + logdet
+            nll = (
+                torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2])
+                - logdet_tot
+            )
+            return nll + logq  # [b]
+        else:
+            flows = list(reversed(self.flows))
+            flows = flows[:-2] + [flows[-1]]  # remove a useless vflow
+            z = (
+                torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype)
+                * noise_scale
+            )
+            for flow in flows:
+                z = flow(z, x_mask, g=x, reverse=reverse)
+            z0, z1 = torch.split(z, [1, 1], 1)
+            logw = z0
+            return logw
+
+
+class DurationPredictor(nn.Module):
+    def __init__(
+        self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = modules.LayerNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = modules.LayerNorm(filter_channels)
+        self.proj = nn.Conv1d(filter_channels, 1, 1)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+    def forward(self, x, x_mask, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
+class TextEncoder(nn.Module):
+    def __init__(
+        self,
+        n_vocab,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        gin_channels=0,
+        num_languages=None,
+        num_tones=None,
+    ):
+        super().__init__()
+        if num_languages is None:
+            from src.text import num_languages
+        if num_tones is None:
+            from src.text import num_tones
+        self.n_vocab = n_vocab
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+        self.emb = nn.Embedding(n_vocab, hidden_channels)
+        nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
+        self.tone_emb = nn.Embedding(num_tones, hidden_channels)
+        nn.init.normal_(self.tone_emb.weight, 0.0, hidden_channels**-0.5)
+        self.language_emb = nn.Embedding(num_languages, hidden_channels)
+        nn.init.normal_(self.language_emb.weight, 0.0, hidden_channels**-0.5)
+        self.bert_proj = nn.Conv1d(1024, hidden_channels, 1)
+        self.ja_bert_proj = nn.Conv1d(768, hidden_channels, 1)
+
+        self.encoder = attentions.Encoder(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, tone, language, bert, ja_bert, g=None):
+        bert_emb = self.bert_proj(bert).transpose(1, 2)
+        ja_bert_emb = self.ja_bert_proj(ja_bert).transpose(1, 2)
+        x = (
+            self.emb(x)
+            + self.tone_emb(tone)
+            + self.language_emb(language)
+            + bert_emb
+            + ja_bert_emb
+        ) * math.sqrt(
+            self.hidden_channels
+        )  # [b, t, h]
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+
+        x = self.encoder(x * x_mask, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return x, m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None, tau=1.0):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * tau * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print("Removing weight norm...")
+        for layer in self.ups:
+            remove_weight_norm(layer)
+        for layer in self.resblocks:
+            layer.remove_weight_norm()
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+            ]
+        )
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm is False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+                norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for layer in self.convs:
+            x = layer(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class ReferenceEncoder(nn.Module):
+    """
+    inputs --- [N, Ty/r, n_mels*r]  mels
+    outputs --- [N, ref_enc_gru_size]
+    """
+
+    def __init__(self, spec_channels, gin_channels=0, layernorm=False):
+        super().__init__()
+        self.spec_channels = spec_channels
+        ref_enc_filters = [32, 32, 64, 64, 128, 128]
+        K = len(ref_enc_filters)
+        filters = [1] + ref_enc_filters
+        convs = [
+            weight_norm(
+                nn.Conv2d(
+                    in_channels=filters[i],
+                    out_channels=filters[i + 1],
+                    kernel_size=(3, 3),
+                    stride=(2, 2),
+                    padding=(1, 1),
+                )
+            )
+            for i in range(K)
+        ]
+        self.convs = nn.ModuleList(convs)
+        # self.wns = nn.ModuleList([weight_norm(num_features=ref_enc_filters[i]) for i in range(K)]) # noqa: E501
+
+        out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K)
+        self.gru = nn.GRU(
+            input_size=ref_enc_filters[-1] * out_channels,
+            hidden_size=256 // 2,
+            batch_first=True,
+        )
+        self.proj = nn.Linear(128, gin_channels)
+        if layernorm:
+            self.layernorm = nn.LayerNorm(self.spec_channels)
+            print('[Ref Enc]: using layer norm')
+        else:
+            self.layernorm = None
+
+    def forward(self, inputs, mask=None):
+        N = inputs.size(0)
+
+        out = inputs.view(N, 1, -1, self.spec_channels)  # [N, 1, Ty, n_freqs]
+        if self.layernorm is not None:
+            out = self.layernorm(out)
+
+        for conv in self.convs:
+            out = conv(out)
+            # out = wn(out)
+            out = F.relu(out)  # [N, 128, Ty//2^K, n_mels//2^K]
+
+        out = out.transpose(1, 2)  # [N, Ty//2^K, 128, n_mels//2^K]
+        T = out.size(1)
+        N = out.size(0)
+        out = out.contiguous().view(N, T, -1)  # [N, Ty//2^K, 128*n_mels//2^K]
+
+        self.gru.flatten_parameters()
+        memory, out = self.gru(out)  # out --- [1, N, 128]
+
+        return self.proj(out.squeeze(0))
+
+    def calculate_channels(self, L, kernel_size, stride, pad, n_convs):
+        for i in range(n_convs):
+            L = (L - kernel_size + 2 * pad) // stride + 1
+        return L
+
+
+class SynthesizerTrn(nn.Module):
+    """
+    Synthesizer for Training
+    """
+
+    def __init__(
+        self,
+        n_vocab,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        n_speakers=256,
+        gin_channels=256,
+        use_sdp=True,
+        n_flow_layer=4,
+        n_layers_trans_flow=6,
+        flow_share_parameter=False,
+        use_transformer_flow=True,
+        use_vc=False,
+        num_languages=None,
+        num_tones=None,
+        norm_refenc=False,
+        **kwargs
+    ):
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.n_speakers = n_speakers
+        self.gin_channels = gin_channels
+        self.n_layers_trans_flow = n_layers_trans_flow
+        self.use_spk_conditioned_encoder = kwargs.get(
+            "use_spk_conditioned_encoder", True
+        )
+        self.use_sdp = use_sdp
+        self.use_noise_scaled_mas = kwargs.get("use_noise_scaled_mas", False)
+        self.mas_noise_scale_initial = kwargs.get("mas_noise_scale_initial", 0.01)
+        self.noise_scale_delta = kwargs.get("noise_scale_delta", 2e-6)
+        self.current_mas_noise_scale = self.mas_noise_scale_initial
+        if self.use_spk_conditioned_encoder and gin_channels > 0:
+            self.enc_gin_channels = gin_channels
+        else:
+            self.enc_gin_channels = 0
+        self.enc_p = TextEncoder(
+            n_vocab,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.enc_gin_channels,
+            num_languages=num_languages,
+            num_tones=num_tones,
+        )
+        self.dec = Generator(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        if use_transformer_flow:
+            self.flow = TransformerCouplingBlock(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers_trans_flow,
+                5,
+                p_dropout,
+                n_flow_layer,
+                gin_channels=gin_channels,
+                share_parameter=flow_share_parameter,
+            )
+        else:
+            self.flow = ResidualCouplingBlock(
+                inter_channels,
+                hidden_channels,
+                5,
+                1,
+                n_flow_layer,
+                gin_channels=gin_channels,
+            )
+        self.sdp = StochasticDurationPredictor(
+            hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels
+        )
+        self.dp = DurationPredictor(
+            hidden_channels, 256, 3, 0.5, gin_channels=gin_channels
+        )
+
+        if n_speakers > 0:
+            self.emb_g = nn.Embedding(n_speakers, gin_channels)
+        else:
+            self.ref_enc = ReferenceEncoder(spec_channels, gin_channels, layernorm=norm_refenc)
+        self.use_vc = use_vc
+
+
+    def forward(self, x, x_lengths, y, y_lengths, sid, tone, language, bert, ja_bert):
+        if self.n_speakers > 0:
+            g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+        else:
+            g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
+        if self.use_vc:
+            g_p = None
+        else:
+            g_p = g
+        x, m_p, logs_p, x_mask = self.enc_p(
+            x, x_lengths, tone, language, bert, ja_bert, g=g_p
+        )
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+
+        with torch.no_grad():
+            # negative cross-entropy
+            s_p_sq_r = torch.exp(-2 * logs_p)  # [b, d, t]
+            neg_cent1 = torch.sum(
+                -0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True
+            )  # [b, 1, t_s]
+            neg_cent2 = torch.matmul(
+                -0.5 * (z_p**2).transpose(1, 2), s_p_sq_r
+            )  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent3 = torch.matmul(
+                z_p.transpose(1, 2), (m_p * s_p_sq_r)
+            )  # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
+            neg_cent4 = torch.sum(
+                -0.5 * (m_p**2) * s_p_sq_r, [1], keepdim=True
+            )  # [b, 1, t_s]
+            neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
+            if self.use_noise_scaled_mas:
+                epsilon = (
+                    torch.std(neg_cent)
+                    * torch.randn_like(neg_cent)
+                    * self.current_mas_noise_scale
+                )
+                neg_cent = neg_cent + epsilon
+
+            attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+            attn = (
+                monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1))
+                .unsqueeze(1)
+                .detach()
+            )
+
+        w = attn.sum(2)
+
+        l_length_sdp = self.sdp(x, x_mask, w, g=g)
+        l_length_sdp = l_length_sdp / torch.sum(x_mask)
+
+        logw_ = torch.log(w + 1e-6) * x_mask
+        logw = self.dp(x, x_mask, g=g)
+        l_length_dp = torch.sum((logw - logw_) ** 2, [1, 2]) / torch.sum(
+            x_mask
+        )  # for averaging
+
+        l_length = l_length_dp + l_length_sdp
+
+        # expand prior
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
+
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        o = self.dec(z_slice, g=g)
+        return (
+            o,
+            l_length,
+            attn,
+            ids_slice,
+            x_mask,
+            y_mask,
+            (z, z_p, m_p, logs_p, m_q, logs_q),
+            (x, logw, logw_),
+        )
+
+    def infer(
+        self,
+        x,
+        x_lengths,
+        sid,
+        tone,
+        language,
+        bert,
+        ja_bert,
+        noise_scale=0.667,
+        length_scale=1,
+        noise_scale_w=0.8,
+        max_len=None,
+        sdp_ratio=0,
+        y=None,
+        g=None,
+    ):
+        # x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, tone, language, bert)
+        # g = self.gst(y)
+        if g is None:
+            if self.n_speakers > 0:
+                g = self.emb_g(sid).unsqueeze(-1)  # [b, h, 1]
+            else:
+                g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
+        if self.use_vc:
+            g_p = None
+        else:
+            g_p = g
+        x, m_p, logs_p, x_mask = self.enc_p(
+            x, x_lengths, tone, language, bert, ja_bert, g=g_p
+        )
+        logw = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w) * (
+            sdp_ratio
+        ) + self.dp(x, x_mask, g=g) * (1 - sdp_ratio)
+        w = torch.exp(logw) * x_mask * length_scale
+        
+        w_ceil = torch.ceil(w)
+        y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
+        y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(
+            x_mask.dtype
+        )
+        attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+        attn = commons.generate_path(w_ceil, attn_mask)
+
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(
+            1, 2
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(
+            1, 2
+        )  # [b, t', t], [b, t, d] -> [b, d, t']
+
+        z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
+        z = self.flow(z_p, y_mask, g=g, reverse=True)
+        o = self.dec((z * y_mask)[:, :, :max_len], g=g)
+        # print('max/min of o:', o.max(), o.min())
+        return o, attn, y_mask, (z, z_p, m_p, logs_p)
+
+    def voice_conversion(self, y, y_lengths, sid_src, sid_tgt, tau=1.0):        
+        g_src = sid_src
+        g_tgt = sid_tgt
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src, tau=tau)
+        z_p = self.flow(z, y_mask, g=g_src)
+        z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
+        o_hat = self.dec(z_hat * y_mask, g=g_tgt)
+        return o_hat, y_mask, (z, z_p, z_hat)

src/nn/__init__.py

@@ -0,0 +1,8 @@
+"""
+Neural network components package
+"""
+
+from .commons import *
+from .attentions import *
+from .modules import *
+from .transforms import *

src/nn/attentions.py

@@ -0,0 +1,459 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from . import commons
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=4,
+        isflow=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.cond_layer_idx = self.n_layers
+        if "gin_channels" in kwargs:
+            self.gin_channels = kwargs["gin_channels"]
+            if self.gin_channels != 0:
+                self.spk_emb_linear = nn.Linear(self.gin_channels, self.hidden_channels)
+                self.cond_layer_idx = (
+                    kwargs["cond_layer_idx"] if "cond_layer_idx" in kwargs else 2
+                )
+                assert (
+                    self.cond_layer_idx < self.n_layers
+                ), "cond_layer_idx should be less than n_layers"
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, g=None):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            if i == self.cond_layer_idx and g is not None:
+                g = self.spk_emb_linear(g.transpose(1, 2))
+                g = g.transpose(1, 2)
+                x = x + g
+                x = x * x_mask
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # pad along column
+        x = F.pad(
+            x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation=None,
+        causal=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+
+        if causal:
+            self.padding = self._causal_padding
+        else:
+            self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+        if self.activation == "gelu":
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+        x = self.conv_2(self.padding(x * x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x

src/nn/commons.py

@@ -0,0 +1,160 @@
+import math
+import torch
+from torch.nn import functional as F
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def convert_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+
+def intersperse(lst, item):
+    result = [item] * (len(lst) * 2 + 1)
+    result[1::2] = lst
+    return result
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    """KL(P||Q)"""
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    """Sample from the Gumbel distribution, protect from overflows."""
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, :, idx_str:idx_end]
+    return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+def convert_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

src/nn/modules.py

@@ -0,0 +1,598 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.nn import Conv1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+from . import commons
+from .commons import init_weights, get_padding
+from .transforms import piecewise_rational_quadratic_transform
+from .attentions import Encoder
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DDSConv(nn.Module):
+    """
+    Dialted and Depth-Separable Convolution
+    """
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super().__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(p_dropout)
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(LayerNorm(channels))
+            self.norms_2.append(LayerNorm(channels))
+
+    def forward(self, x, x_mask, g=None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(self, x, x_mask, g=None, **kwargs):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i in range(self.n_layers):
+            x_in = self.in_layers[i](x)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = self.res_skip_layers[i](acts)
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Log(nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class Flip(nn.Module):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=p_dropout,
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x
+
+
+class TransformerCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        n_layers,
+        n_heads,
+        p_dropout=0,
+        filter_channels=0,
+        mean_only=False,
+        wn_sharing_parameter=None,
+        gin_channels=0,
+    ):
+        assert n_layers == 3, n_layers
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = (
+            Encoder(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=gin_channels,
+            )
+            if wn_sharing_parameter is None
+            else wn_sharing_parameter
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x

src/nn/transforms.py

@@ -0,0 +1,209 @@
+import torch
+from torch.nn import functional as F
+
+import numpy as np
+
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

src/text/__init__.py

@@ -0,0 +1,24 @@
+from .symbols import *
+
+
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+
+
+def cleaned_text_to_sequence(cleaned_text, tones, language, symbol_to_id=None):
+    """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+    Returns:
+      List of integers corresponding to the symbols in the text
+    """
+    symbol_to_id_map = symbol_to_id if symbol_to_id else _symbol_to_id
+    unk_id = symbol_to_id_map.get("UNK")
+    if unk_id is None:
+        phones = [symbol_to_id_map[symbol] for symbol in cleaned_text]
+    else:
+        phones = [symbol_to_id_map.get(symbol, unk_id) for symbol in cleaned_text]
+    tone_start = language_tone_start_map[language]
+    tones = [i + tone_start for i in tones]
+    lang_id = language_id_map[language]
+    lang_ids = [lang_id for _ in phones]
+    return phones, tones, lang_ids

src/text/cleaner.py

@@ -0,0 +1,44 @@
+from . import cleaned_text_to_sequence
+import copy
+
+_language_modules = {}
+
+def _get_language_module(language):
+    """Lazy import language modules to avoid unnecessary dependencies."""
+    if language == 'VI':
+        from . import vietnamese
+        _language_modules['VI'] = vietnamese
+    else:
+        raise ValueError(f"Unsupported language: {language}")
+    
+    return _language_modules[language]
+
+
+def clean_text(text, language):
+    language_module = _get_language_module(language)
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    return norm_text, phones, tones, word2ph
+
+
+def clean_text_bert(text, language, device=None):
+    language_module = _get_language_module(language)
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    
+    word2ph_bak = copy.deepcopy(word2ph)
+    for i in range(len(word2ph)):
+        word2ph[i] = word2ph[i] * 2
+    word2ph[0] += 1
+    bert = language_module.get_bert_feature(norm_text, word2ph, device=device)
+    
+    return norm_text, phones, tones, word2ph_bak, bert
+
+
+def text_to_sequence(text, language):
+    norm_text, phones, tones, word2ph = clean_text(text, language)
+    return cleaned_text_to_sequence(phones, tones, language)
+
+
+if __name__ == "__main__":
+    pass

src/text/symbols.py

@@ -0,0 +1,373 @@
+# punctuation = ["!", "?", "…", ",", ".", "'", "-"]
+punctuation = ["!", "?", "…", ",", ".", "'", "-", "¿", "¡"]
+pu_symbols = punctuation + ["SP", "UNK"]
+pad = "_"
+
+# chinese
+zh_symbols = [
+    "E",
+    "En",
+    "a",
+    "ai",
+    "an",
+    "ang",
+    "ao",
+    "b",
+    "c",
+    "ch",
+    "d",
+    "e",
+    "ei",
+    "en",
+    "eng",
+    "er",
+    "f",
+    "g",
+    "h",
+    "i",
+    "i0",
+    "ia",
+    "ian",
+    "iang",
+    "iao",
+    "ie",
+    "in",
+    "ing",
+    "iong",
+    "ir",
+    "iu",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "ong",
+    "ou",
+    "p",
+    "q",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "u",
+    "ua",
+    "uai",
+    "uan",
+    "uang",
+    "ui",
+    "un",
+    "uo",
+    "v",
+    "van",
+    "ve",
+    "vn",
+    "w",
+    "x",
+    "y",
+    "z",
+    "zh",
+    "AA",
+    "EE",
+    "OO",
+]
+num_zh_tones = 6
+
+# japanese
+ja_symbols = [
+    "N",
+    "a",
+    "a:",
+    "b",
+    "by",
+    "ch",
+    "d",
+    "dy",
+    "e",
+    "e:",
+    "f",
+    "g",
+    "gy",
+    "h",
+    "hy",
+    "i",
+    "i:",
+    "j",
+    "k",
+    "ky",
+    "m",
+    "my",
+    "n",
+    "ny",
+    "o",
+    "o:",
+    "p",
+    "py",
+    "q",
+    "r",
+    "ry",
+    "s",
+    "sh",
+    "t",
+    "ts",
+    "ty",
+    "u",
+    "u:",
+    "w",
+    "y",
+    "z",
+    "zy",
+]
+num_ja_tones = 1
+
+# English
+en_symbols = [
+    "aa",
+    "ae",
+    "ah",
+    "ao",
+    "aw",
+    "ay",
+    "b",
+    "ch",
+    "d",
+    "dh",
+    "eh",
+    "er",
+    "ey",
+    "f",
+    "g",
+    "hh",
+    "ih",
+    "iy",
+    "jh",
+    "k",
+    "l",
+    "m",
+    "n",
+    "ng",
+    "ow",
+    "oy",
+    "p",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "th",
+    "uh",
+    "uw",
+    "V",
+    "w",
+    "y",
+    "z",
+    "zh",
+]
+num_en_tones = 4
+
+# Korean
+kr_symbols = ['ᄌ', 'ᅥ', 'ᆫ', 'ᅦ', 'ᄋ', 'ᅵ', 'ᄅ', 'ᅴ', 'ᄀ', 'ᅡ', 'ᄎ', 'ᅪ', 'ᄑ', 'ᅩ', 'ᄐ', 'ᄃ', 'ᅢ', 'ᅮ', 'ᆼ', 'ᅳ', 'ᄒ', 'ᄆ', 'ᆯ', 'ᆷ', 'ᄂ', 'ᄇ', 'ᄉ', 'ᆮ', 'ᄁ', 'ᅬ', 'ᅣ', 'ᄄ', 'ᆨ', 'ᄍ', 'ᅧ', 'ᄏ', 'ᆸ', 'ᅭ', '(', 'ᄊ', ')', 'ᅲ', 'ᅨ', 'ᄈ', 'ᅱ', 'ᅯ', 'ᅫ', 'ᅰ', 'ᅤ', '~', '\\', '[', ']', '/', '^', ':', 'ㄸ', '*']
+num_kr_tones = 1
+
+# Spanish
+es_symbols = [
+        "N",
+        "Q",
+        "a",
+        "b",
+        "d",
+        "e",
+        "f",
+        "g",
+        "h",
+        "i",
+        "j",
+        "k",
+        "l",
+        "m",
+        "n",
+        "o",
+        "p",
+        "s",
+        "t",
+        "u",
+        "v",
+        "w",
+        "x",
+        "y",
+        "z",
+        "ɑ",
+        "æ",
+        "ʃ",
+        "ʑ",
+        "ç",
+        "ɯ",
+        "ɪ",
+        "ɔ",
+        "ɛ",
+        "ɹ",
+        "ð",
+        "ə",
+        "ɫ",
+        "ɥ",
+        "ɸ",
+        "ʊ",
+        "ɾ",
+        "ʒ",
+        "θ",
+        "β",
+        "ŋ",
+        "ɦ",
+        "ɡ",
+        "r",
+        "ɲ",
+        "ʝ",
+        "ɣ",
+        "ʎ",
+        "ˈ",
+        "ˌ",
+        "ː"
+    ]
+num_es_tones = 1
+
+# French 
+fr_symbols = [
+    "\u0303",
+    "œ",
+    "ø",
+    "ʁ",
+    "ɒ",
+    "ʌ",
+    "ɜ",
+    "ɐ"
+]
+num_fr_tones = 1
+
+# German 
+de_symbols = [
+    "ʏ",
+    "̩"
+  ]
+num_de_tones = 1
+
+# Russian 
+ru_symbols = [
+    "ɭ",
+    "ʲ",
+    "ɕ",
+    "\"",
+    "ɵ",
+    "^",
+    "ɬ"
+]
+num_ru_tones = 1
+
+# Vietnamese (IPA-based, compatible with VieNeu-TTS-140h dataset)
+vi_symbols = [
+    # Consonants (simple)
+    "ʈ",   # tr
+    "ɖ",   # đ
+    "ɗ",   # implosive d (đ variant)
+    "ɓ",   # implosive b
+    "ʰ",   # aspiration marker
+    "ă",   # short a (Vietnamese)
+    "ʷ",   # labialization marker
+    "̆",    # breve diacritic
+    "͡",    # tie bar (for affricates)
+    "ʤ",   # voiced postalveolar affricate
+    "ʧ",   # voiceless postalveolar affricate
+    # Foreign/special characters found in dataset
+    "т",   # Cyrillic т
+    "輪",  # Chinese character
+    "и",   # Cyrillic и
+    "л",   # Cyrillic л
+    "р",   # Cyrillic р
+    "µ",   # micro sign
+    "ʂ",   # s (retroflex)
+    "ʐ",   # r (retroflex)
+    "ʔ",   # glottal stop
+    "ɣ",   # g (southern)
+    # Multi-char consonants (from vietnamese.py g2p)
+    "tʰ",  # th
+    "kʰ",  # kh
+    "kw",  # qu -> kw
+    "tʃ",  # ch
+    "ɹ",   # r IPA
+    # Vowels specific to Vietnamese
+    "ɤ",   # ơ
+    "ɐ",   # a short
+    "ɑ",   # a back
+    "ɨ",   # ư variant
+    "ʉ",   # u variant
+    "ɜ",   # open-mid central
+    # Long vowels (from VieNeu-TTS dataset)
+    "əː",  # schwa long
+    "aː",  # a long  
+    "ɜː",  # open-mid central long
+    "ɑː",  # open back long
+    "ɔː",  # open-mid back long
+    "iː",  # close front long
+    "uː",  # close back long
+    "eː",  # close-mid front long
+    "oː",  # close-mid back long
+    # Diphthongs and special combinations
+    "iə",  # ia/iê
+    "ɨə",  # ưa/ươ
+    "uə",  # ua/uô
+    # Additional IPA markers
+    "ˑ",   # half-long
+    "̪",    # dental diacritic
+    # Tone-related (though tones are handled separately)
+    "˥",   # tone 1 marker
+    "˩",   # tone marker
+    "˧",   # tone marker
+    "˨",   # tone marker
+    "˦",   # tone marker
+    # Numbers (found in phonemized dataset)
+    "0", "1", "2", "3", "4", "5", "6", "7", "8", "9",
+    # Special characters from dataset
+    "$", "%", "&", "«", "»", "–", "ı",
+    # viphoneme specific symbols
+    "wʷ",  # labialized w
+    "#",   # unknown/fallback marker
+    "ô",   # Vietnamese ô (fallback)
+    "ʃ",   # voiceless postalveolar fricative
+    "ʒ",   # voiced postalveolar fricative
+    "θ",   # voiceless dental fricative
+    "ð",   # voiced dental fricative
+    "æ",   # near-open front unrounded
+    "ɪ",   # near-close front unrounded
+    "ʊ",   # near-close back rounded 
+    # Vietnamese fallback characters (when viphoneme fails to parse)
+    "ẩ", "ò", "à", "á", "ủ", "ờ", "ộ", "ả", "ó", "é", "ê",
+    "ồ", "ấ", "ú", "ế", "ớ", "ì", "ọ", "ố", "ư", "ữ",
+]
+num_vi_tones = 8  # 6 tones + 1 neutral + 1 extra for data compatibility
+
+# combine all symbols
+normal_symbols = sorted(set(zh_symbols + ja_symbols + en_symbols + kr_symbols + es_symbols + fr_symbols + de_symbols + ru_symbols + vi_symbols))
+symbols = [pad] + normal_symbols + pu_symbols
+sil_phonemes_ids = [symbols.index(i) for i in pu_symbols]
+
+# combine all tones
+num_tones = num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones + num_ru_tones + num_vi_tones
+
+# language maps
+language_id_map = {"ZH": 0, "JP": 1, "EN": 2, "ZH_MIX_EN": 3, 'KR': 4, 'ES': 5, 'SP': 5, 'FR': 6, 'VI': 7}
+num_languages = len(language_id_map.keys())
+
+language_tone_start_map = {
+    "ZH": 0,
+    "ZH_MIX_EN": 0,
+    "JP": num_zh_tones,
+    "EN": num_zh_tones + num_ja_tones,
+    'KR': num_zh_tones + num_ja_tones + num_en_tones,
+    "ES": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "SP": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "FR": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones,
+    "VI": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones + num_ru_tones,
+}
+
+if __name__ == "__main__":
+    a = set(zh_symbols)
+    b = set(en_symbols)
+    print(sorted(a & b))

src/text/vietnamese.py

@@ -0,0 +1,429 @@
+import re
+import unicodedata
+from transformers import AutoTokenizer
+from . import punctuation, symbols
+
+# Vietnamese BERT model
+model_id = 'vinai/phobert-base-v2'
+tokenizer = None
+
+def get_tokenizer():
+    global tokenizer
+    if tokenizer is None:
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+    return tokenizer
+
+# Vietnamese IPA phoneme set based on VieNeu-TTS-140h dataset
+# These are extracted from the phonemized_text field in the dataset
+VI_IPA_CONSONANTS = [
+    'b', 'c', 'd', 'f', 'g', 'h', 'j', 'k', 'l', 'm', 'n', 'p', 'r', 's', 't', 'v', 'w', 'x', 'z',
+    'ŋ',  # ng
+    'ɲ',  # nh
+    'ʈ',  # tr
+    'ɖ',  # đ
+    'tʰ', # th
+    'kʰ', # kh
+    'ʂ',  # s (southern)
+    'ɣ',  # g (southern)
+    'χ',  # x (some dialects)
+]
+
+VI_IPA_VOWELS = [
+    'a', 'ă', 'â', 'e', 'ê', 'i', 'o', 'ô', 'ơ', 'u', 'ư', 'y',
+    'ə',  # ơ
+    'ɛ',  # e
+    'ɔ',  # o
+    'ɯ',  # ư
+    'ɤ',  # ơ variant
+    'ɐ',  # a short
+    'ʊ',  # u short
+    'ɪ',  # i short
+    'ʌ',  # â
+    'æ',  # a variant
+]
+
+# Vietnamese tone markers (numbers 1-6 or ˈ ˌ for stress)
+VI_TONE_MARKERS = ['1', '2', '3', '4', '5', '6', 'ˈ', 'ˌ', 'ː']
+
+# Combined IPA symbols used in VieNeu-TTS dataset
+VI_IPA_SYMBOLS = [
+    # Consonants
+    'b', 'c', 'd', 'f', 'g', 'h', 'j', 'k', 'l', 'm', 'n', 'p', 'r', 's', 't', 'v', 'w', 'x', 'z',
+    'ŋ', 'ɲ', 'ʈ', 'ɖ', 'ʂ', 'ɣ', 'χ', 'ʔ',
+    # Vowels
+    'a', 'ă', 'e', 'i', 'o', 'u', 'y',
+    'ə', 'ɛ', 'ɔ', 'ɯ', 'ɤ', 'ɐ', 'ʊ', 'ɪ', 'ʌ', 'æ', 'ɑ',
+    # Special markers
+    'ˈ', 'ˌ', 'ː',
+    # Tone numbers
+    '1', '2', '3', '4', '5', '6',
+]
+
+def normalize_vietnamese_text(text):
+    """Normalize Vietnamese text."""
+    # Normalize unicode
+    text = unicodedata.normalize('NFC', text)
+    
+    # Remove extra whitespace
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    
+    # Convert numbers to words (basic)
+    text = convert_numbers_to_vietnamese(text)
+    
+    return text
+
+def convert_numbers_to_vietnamese(text):
+    """Convert numbers to Vietnamese words (basic implementation)."""
+    num_map = {
+        '0': 'không', '1': 'một', '2': 'hai', '3': 'ba', '4': 'bốn',
+        '5': 'năm', '6': 'sáu', '7': 'bảy', '8': 'tám', '9': 'chín',
+        '10': 'mười', '100': 'trăm', '1000': 'nghìn'
+    }
+    
+    # Simple replacement for single digits in context
+    def replace_num(match):
+        num = match.group(0)
+        if num in num_map:
+            return num_map[num]
+        return num
+    
+    # Only replace standalone numbers
+    text = re.sub(r'\b\d\b', replace_num, text)
+    return text
+
+def text_normalize(text):
+    """Normalize text for Vietnamese TTS."""
+    text = normalize_vietnamese_text(text)
+    return text
+
+def parse_ipa_phonemes(phonemized_text):
+    """
+    Parse IPA phonemized text from VieNeu-TTS dataset.
+    Example: "ŋˈyə2j ŋˈyə2j bˈan xwˈan vˈe2"
+    Returns: phones, tones, word2ph
+    """
+    phones = []
+    tones = []
+    word2ph = []
+    
+    # Split by space to get words
+    words = phonemized_text.strip().split()
+    
+    for word in words:
+        word_phones = []
+        word_tones = []
+        
+        # Parse each character/symbol in the word
+        i = 0
+        current_tone = 0  # Default tone (neutral/tone 1)
+        
+        while i < len(word):
+            char = word[i]
+            
+            # Check for tone numbers (1-6)
+            if char.isdigit():
+                current_tone = int(char)
+                i += 1
+                continue
+            
+            # Check for stress markers
+            if char in ['ˈ', 'ˌ']:
+                # Primary or secondary stress - could be used as tone variant
+                i += 1
+                continue
+            
+            # Check for length marker
+            if char == 'ː':
+                # Long vowel marker - append to previous phone if exists
+                if word_phones:
+                    word_phones[-1] = word_phones[-1] + 'ː'
+                i += 1
+                continue
+            
+            # Check for punctuation
+            if char in punctuation:
+                if word_phones:
+                    phones.extend(word_phones)
+                    tones.extend([current_tone] * len(word_phones))
+                    word2ph.append(len(word_phones))
+                    word_phones = []
+                    word_tones = []
+                phones.append(char)
+                tones.append(0)
+                word2ph.append(1)
+                i += 1
+                continue
+            
+            # Regular phoneme
+            word_phones.append(char)
+            i += 1
+        
+        # Apply collected tone to all phones in this word
+        if word_phones:
+            phones.extend(word_phones)
+            tones.extend([current_tone] * len(word_phones))
+            word2ph.append(len(word_phones))
+    
+    return phones, tones, word2ph
+
+def g2p_ipa(text):
+    """
+    Convert text to phonemes using external IPA converter.
+    This is a fallback for when phonemized_text is not available.
+    For training, we use the pre-phonemized text from the dataset.
+    """
+    try:
+        from viphoneme import vi2ipa
+        phonemized = vi2ipa(text)
+        phones, tones, word2ph = parse_ipa_phonemes(phonemized)
+    except ImportError:
+        # Fallback: use character-based representation
+        phones, tones, word2ph = g2p_char_based(text)
+    
+    # Add start and end tokens
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    
+    return phones, tones, word2ph
+
+def g2p_char_based(text):
+    """
+    Character-based G2P with Vietnamese to IPA mapping.
+    """
+    phones = []
+    tones = []
+    word2ph = []
+    
+    # Vietnamese tone marks to tone number mapping
+    tone_marks = {
+        '\u0300': 2,  # à - huyền
+        '\u0301': 1,  # á - sắc  
+        '\u0303': 3,  # ã - ngã
+        '\u0309': 4,  # ả - hỏi
+        '\u0323': 5,  # ạ - nặng
+    }
+    
+    # Vietnamese character to IPA mapping (COMPREHENSIVE - matching training data)
+    # Multi-char outputs are split into lists to avoid KeyError for missing multi-char symbols
+    vi_to_ipa = {
+        # Multi-char consonants (check these first - ORDER MATTERS)
+        'ngh': 'ŋ',
+        'ng': 'ŋ',
+        'nh': 'ɲ',
+        'ch': ['t', 'ʃ'],  # Vietnamese ch = IPA t + ʃ (separated in training data)
+        'tr': 'ʈ',   # retroflex
+        'th': ['t', 'h'],   # aspirated th
+        'ph': 'f',
+        'kh': 'x',   # Vietnamese 'kh' = IPA 'x' (matches training data)
+        'gh': 'ɣ',
+        'gi': 'z',
+        'qu': 'kw',   # qu -> kw (single symbol in training data)
+        # Special Vietnamese consonants
+        'đ': 'ɗ',    # implosive d
+        # Basic consonants that need IPA mapping
+        'x': 's',    # Vietnamese 'x' = IPA 's'
+        'c': 'k',    # Vietnamese 'c' = IPA 'k'
+        'd': 'z',    # Vietnamese 'd' (northern) = 'z'
+        'r': 'ɹ',    # Vietnamese 'r' = IPA 'ɹ' (matches training data)
+        's': 's',
+        'b': 'b',
+        'g': 'ɣ',
+        'h': 'h',
+        'k': 'k',
+        'l': 'l',
+        'm': 'm',
+        'n': 'n',
+        'p': 'p',
+        't': 't',
+        'v': 'v',
+        'f': 'f',
+        'j': 'j',
+        'w': 'w',
+        'y': 'j',    # Vietnamese 'y' = IPA 'j' (matches training data)
+        # Vowels - MUST match training data phonemes exactly!
+        'a': 'aː',   # Long 'a' (matches training: aː)
+        'ă': 'a',    # Short 'a' 
+        'â': 'ə',    # schwa
+        'e': 'ɛ',    # open-mid (matches training: ɛ)
+        'ê': 'e',    # close-mid
+        'i': 'i',
+        'o': 'ɔ',    # open-mid back (matches training: ɔ)
+        'ô': 'o',    # close-mid back
+        'ơ': 'əː',   # long schwa
+        'u': 'u',
+        'ư': 'ɯ',    # close back unrounded
+    }
+    
+    words = text.split()
+    for word in words:
+        # Decompose to separate base char and tone mark
+        decomposed = unicodedata.normalize('NFD', word)
+        word_phones = []
+        current_tone = 0
+        
+        i = 0
+        chars = list(decomposed)
+        while i < len(chars):
+            char = chars[i]
+            
+            if char in tone_marks:
+                current_tone = tone_marks[char]
+                i += 1
+                continue
+            
+            if char in punctuation:
+                if word_phones:
+                    phones.extend(word_phones)
+                    tones.extend([current_tone] * len(word_phones))
+                    word2ph.append(len(word_phones))
+                    word_phones = []
+                phones.append(char)
+                tones.append(0)
+                word2ph.append(1)
+                current_tone = 0
+                i += 1
+                continue
+            
+            if unicodedata.combining(char):
+                i += 1
+                continue
+            
+            # Check for multi-char sequences (digraphs/trigraphs)
+            lower_char = char.lower()
+            matched = False
+            
+            # Try trigraphs first
+            if i + 2 < len(chars):
+                trigraph = (lower_char + chars[i+1].lower() + chars[i+2].lower())
+                if trigraph in vi_to_ipa:
+                    result = vi_to_ipa[trigraph]
+                    if isinstance(result, list):
+                        word_phones.extend(result)
+                    else:
+                        word_phones.append(result)
+                    i += 3
+                    matched = True
+            
+            # Try digraphs
+            if not matched and i + 1 < len(chars):
+                digraph = lower_char + chars[i+1].lower()
+                if digraph in vi_to_ipa:
+                    result = vi_to_ipa[digraph]
+                    if isinstance(result, list):
+                        word_phones.extend(result)
+                    else:
+                        word_phones.append(result)
+                    i += 2
+                    matched = True
+            
+            # Single char
+            if not matched:
+                if lower_char in vi_to_ipa:
+                    result = vi_to_ipa[lower_char]
+                    if isinstance(result, list):
+                        word_phones.extend(result)
+                    else:
+                        word_phones.append(result)
+                else:
+                    word_phones.append(lower_char)
+                i += 1
+        
+        if word_phones:
+            phones.extend(word_phones)
+            tones.extend([current_tone] * len(word_phones))
+            word2ph.append(len(word_phones))
+    
+    # Add boundary tokens
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    
+    return phones, tones, word2ph
+
+def g2p(text):
+    """
+    Main G2P function for Vietnamese.
+    Uses character-to-IPA mapping with BERT alignment.
+    """
+    tok = get_tokenizer()
+    norm_text = text_normalize(text)
+    
+    # Tokenize for BERT alignment
+    tokenized = tok.tokenize(norm_text)
+    
+    # Use character-based G2P with IPA mapping
+    phones, tones, word2ph = g2p_char_based(norm_text)
+    
+    # Ensure word2ph aligns with tokenized output
+    # PhoBERT uses subword tokenization, so we need to distribute phones
+    if len(word2ph) != len(tokenized) + 2:  # +2 for start/end tokens
+        # Redistribute word2ph to match tokenized length
+        total_phones = sum(word2ph)
+        new_word2ph = distribute_phones(total_phones, len(tokenized))
+        word2ph = [1] + new_word2ph + [1]
+    
+    return phones, tones, word2ph
+
+def g2p_with_phonemes(text, phonemized_text):
+    """
+    G2P using pre-phonemized text from dataset.
+    This is the recommended method for training.
+    """
+    tok = get_tokenizer()
+    
+    # Parse IPA phonemes
+    phones, tones, word2ph = parse_ipa_phonemes(phonemized_text)
+    
+    # Add boundary tokens
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    
+    # Get tokenized text for BERT alignment
+    tokenized = tok.tokenize(text)
+    
+    # Distribute word2ph to match tokenized output + boundaries
+    if word2ph:
+        total_phones = sum(word2ph)
+        new_word2ph = distribute_phones(total_phones, len(tokenized))
+        word2ph = [1] + new_word2ph + [1]
+    else:
+        word2ph = [1] + [1] * len(tokenized) + [1]
+    
+    return phones, tones, word2ph
+
+def distribute_phones(n_phone, n_word):
+    """Distribute phones across words as evenly as possible."""
+    if n_word == 0:
+        return []
+    phones_per_word = [n_phone // n_word] * n_word
+    remainder = n_phone % n_word
+    for i in range(remainder):
+        phones_per_word[i] += 1
+    return phones_per_word
+
+def get_bert_feature(text, word2ph, device='cuda'):
+    """Get BERT features for Vietnamese text."""
+    from . import vietnamese_bert
+    return vietnamese_bert.get_bert_feature(text, word2ph, device=device, model_id=model_id)
+
+
+if __name__ == "__main__":
+    # Test
+    test_text = "Xin chào, tôi là một trợ lý AI."
+    test_phonemes = "sˈin tʂˈaːw, tˈoj lˈaː2 mˈo6t tʂˈɤ4 lˈi4 ˌaːˈi."
+    
+    print("Test text:", test_text)
+    print("Normalized:", text_normalize(test_text))
+    
+    # Test with phonemes
+    phones, tones, word2ph = g2p_with_phonemes(test_text, test_phonemes)
+    print("Phones:", phones)
+    print("Tones:", tones)
+    print("Word2Ph:", word2ph)
+    
+    # Test without phonemes
+    phones2, tones2, word2ph2 = g2p(test_text)
+    print("\nChar-based phones:", phones2)
+    print("Char-based tones:", tones2)

src/utils/__init__.py

@@ -0,0 +1,5 @@
+"""
+Utility functions package
+"""
+
+from .helpers import *

src/utils/helpers.py

@@ -0,0 +1,452 @@
+import os
+import glob
+import argparse
+import logging
+import json
+import subprocess
+import numpy as np
+from scipy.io.wavfile import read
+import torch
+import torchaudio
+import librosa
+from src.text import cleaned_text_to_sequence
+from src.text.cleaner import clean_text
+from src.nn import commons
+
+MATPLOTLIB_FLAG = False
+
+logger = logging.getLogger(__name__)
+
+
+
+def get_text_for_tts_infer(text, language_str, hps, device, symbol_to_id=None):
+    norm_text, phone, tone, word2ph = clean_text(text, language_str)
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str, symbol_to_id)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+
+    if getattr(hps.data, "disable_bert", False):
+        bert = torch.zeros(1024, len(phone))
+        ja_bert = torch.zeros(768, len(phone))
+    else:
+        bert = get_bert(norm_text, word2ph, language_str, device)
+        del word2ph
+        assert bert.shape[-1] == len(phone), phone
+
+        if language_str == "ZH":
+            bert = bert
+            ja_bert = torch.zeros(768, len(phone))
+        elif language_str in ["JP", "EN", "ZH_MIX_EN", 'KR', 'SP', 'ES', 'FR', 'DE', 'RU', 'VI']:
+            ja_bert = bert
+            bert = torch.zeros(1024, len(phone))
+        else:
+            raise NotImplementedError()
+
+    assert bert.shape[-1] == len(
+        phone
+    ), f"Bert seq len {bert.shape[-1]} != {len(phone)}"
+
+    phone = torch.LongTensor(phone)
+    tone = torch.LongTensor(tone)
+    language = torch.LongTensor(language)
+    return bert, ja_bert, phone, tone, language
+
+def load_checkpoint(checkpoint_path, model, optimizer=None, skip_optimizer=False):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+    iteration = checkpoint_dict.get("iteration", 0)
+    learning_rate = checkpoint_dict.get("learning_rate", 0.)
+    if (
+        optimizer is not None
+        and not skip_optimizer
+        and checkpoint_dict["optimizer"] is not None
+    ):
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    elif optimizer is None and not skip_optimizer:
+        # else:      Disable this line if Infer and resume checkpoint,then enable the line upper
+        new_opt_dict = optimizer.state_dict()
+        new_opt_dict_params = new_opt_dict["param_groups"][0]["params"]
+        new_opt_dict["param_groups"] = checkpoint_dict["optimizer"]["param_groups"]
+        new_opt_dict["param_groups"][0]["params"] = new_opt_dict_params
+        optimizer.load_state_dict(new_opt_dict)
+
+    saved_state_dict = checkpoint_dict["model"]
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+
+    new_state_dict = {}
+    for k, v in state_dict.items():
+        try:
+            # assert "emb_g" not in k
+            new_state_dict[k] = saved_state_dict[k]
+            assert saved_state_dict[k].shape == v.shape, (
+                saved_state_dict[k].shape,
+                v.shape,
+            )
+        except Exception as e:
+            print(e)
+            # For upgrading from the old version
+            if "ja_bert_proj" in k:
+                v = torch.zeros_like(v)
+                logger.warn(
+                    f"Seems you are using the old version of the model, the {k} is automatically set to zero for backward compatibility"
+                )
+            else:
+                logger.error(f"{k} is not in the checkpoint")
+
+            new_state_dict[k] = v
+
+    if hasattr(model, "module"):
+        model.module.load_state_dict(new_state_dict, strict=False)
+    else:
+        model.load_state_dict(new_state_dict, strict=False)
+
+    logger.info(
+        "Loaded checkpoint '{}' (iteration {})".format(checkpoint_path, iteration)
+    )
+
+    return model, optimizer, learning_rate, iteration
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at iteration {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    torch.save(
+        {
+            "model": state_dict,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def summarize(
+    writer,
+    global_step,
+    scalars={},
+    histograms={},
+    images={},
+    audios={},
+    audio_sampling_rate=22050,
+):
+    for k, v in scalars.items():
+        writer.add_scalar(k, v, global_step)
+    for k, v in histograms.items():
+        writer.add_histogram(k, v, global_step)
+    for k, v in images.items():
+        writer.add_image(k, v, global_step, dataformats="HWC")
+    for k, v in audios.items():
+        writer.add_audio(k, v, global_step, audio_sampling_rate)
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+    f_list = glob.glob(os.path.join(dir_path, regex))
+    f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
+    x = f_list[-1]
+    return x
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        try:
+            import matplotlib
+
+            matplotlib.use("Agg")
+            MATPLOTLIB_FLAG = True
+            mpl_logger = logging.getLogger("matplotlib")
+            mpl_logger.setLevel(logging.WARNING)
+        except Exception:
+            spec = np.asarray(spectrogram, dtype=np.float32)
+            if spec.ndim > 2:
+                spec = np.squeeze(spec)
+            if spec.ndim != 2:
+                return np.zeros((1, 1, 3), dtype=np.uint8)
+            vmin = np.nanmin(spec)
+            vmax = np.nanmax(spec)
+            if not np.isfinite(vmin) or not np.isfinite(vmax) or vmax <= vmin:
+                return np.zeros((spec.shape[0], spec.shape[1], 3), dtype=np.uint8)
+            img = ((spec - vmin) / (vmax - vmin) * 255.0).clip(0, 255).astype(np.uint8)
+            return np.stack([img, img, img], axis=-1)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    # Use buffer_rgba() instead of deprecated tostring_rgb()
+    buf = fig.canvas.buffer_rgba()
+    data = np.asarray(buf)[:, :, :3]  # Remove alpha channel
+    plt.close()
+    return data
+
+
+def plot_alignment_to_numpy(alignment, info=None):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        try:
+            import matplotlib
+
+            matplotlib.use("Agg")
+            MATPLOTLIB_FLAG = True
+            mpl_logger = logging.getLogger("matplotlib")
+            mpl_logger.setLevel(logging.WARNING)
+        except Exception:
+            ali = np.asarray(alignment, dtype=np.float32)
+            if ali.ndim > 2:
+                ali = np.squeeze(ali)
+            if ali.ndim != 2:
+                return np.zeros((1, 1, 3), dtype=np.uint8)
+            vmin = np.nanmin(ali)
+            vmax = np.nanmax(ali)
+            if not np.isfinite(vmin) or not np.isfinite(vmax) or vmax <= vmin:
+                return np.zeros((ali.shape[0], ali.shape[1], 3), dtype=np.uint8)
+            img = ((ali - vmin) / (vmax - vmin) * 255.0).clip(0, 255).astype(np.uint8)
+            return np.stack([img, img, img], axis=-1)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(6, 4))
+    im = ax.imshow(
+        alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
+    )
+    fig.colorbar(im, ax=ax)
+    xlabel = "Decoder timestep"
+    if info is not None:
+        xlabel += "\n\n" + info
+    plt.xlabel(xlabel)
+    plt.ylabel("Encoder timestep")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    # Use buffer_rgba() instead of deprecated tostring_rgb()
+    buf = fig.canvas.buffer_rgba()
+    data = np.asarray(buf)[:, :, :3]  # Remove alpha channel
+    plt.close()
+    return data
+
+
+def load_wav_to_torch(full_path):
+    sampling_rate, data = read(full_path)
+    return torch.FloatTensor(data.astype(np.float32)), sampling_rate
+
+
+def load_wav_to_torch_new(full_path):
+    audio_norm, sampling_rate = torchaudio.load(full_path, frame_offset=0, num_frames=-1, normalize=True, channels_first=True)
+    audio_norm = audio_norm.mean(dim=0)
+    return audio_norm, sampling_rate
+
+def load_wav_to_torch_librosa(full_path, sr):
+    audio_norm, sampling_rate = librosa.load(full_path, sr=sr, mono=True)
+    return torch.FloatTensor(audio_norm.astype(np.float32)), sampling_rate
+
+
+def load_filepaths_and_text(filename, split="|"):
+    with open(filename, encoding="utf-8") as f:
+        filepaths_and_text = [line.strip().split(split) for line in f]
+    return filepaths_and_text
+
+
+def get_hparams(init=True):
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-c",
+        "--config",
+        type=str,
+        default="./configs/base.json",
+        help="JSON file for configuration",
+    )
+    parser.add_argument('--local_rank', type=int, default=0)
+    parser.add_argument('--world-size', type=int, default=1)
+    parser.add_argument('--port', type=int, default=10000)
+    parser.add_argument("-m", "--model", type=str, required=True, help="Model name")
+    parser.add_argument('--pretrain_G', type=str, default=None,
+                            help='pretrain model')
+    parser.add_argument('--pretrain_D', type=str, default=None,
+                            help='pretrain model D')
+    parser.add_argument('--pretrain_dur', type=str, default=None,
+                            help='pretrain model duration')
+
+    args = parser.parse_args()
+    model_dir = os.path.join("./logs", args.model)
+
+    os.makedirs(model_dir, exist_ok=True)
+
+    config_path = args.config
+    config_save_path = os.path.join(model_dir, "config.json")
+    if init:
+        with open(config_path, "r") as f:
+            data = f.read()
+        with open(config_save_path, "w") as f:
+            f.write(data)
+    else:
+        with open(config_save_path, "r") as f:
+            data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    hparams.pretrain_G = args.pretrain_G
+    hparams.pretrain_D = args.pretrain_D
+    hparams.pretrain_dur = args.pretrain_dur
+    hparams.port = args.port
+    return hparams
+
+
+def clean_checkpoints(path_to_models="logs/44k/", n_ckpts_to_keep=2, sort_by_time=True):
+    """Freeing up space by deleting saved ckpts
+
+    Arguments:
+    path_to_models    --  Path to the model directory
+    n_ckpts_to_keep   --  Number of ckpts to keep, excluding G_0.pth and D_0.pth
+    sort_by_time      --  True -> chronologically delete ckpts
+                          False -> lexicographically delete ckpts
+    """
+    import re
+
+    ckpts_files = [
+        f
+        for f in os.listdir(path_to_models)
+        if os.path.isfile(os.path.join(path_to_models, f))
+    ]
+
+    def name_key(_f):
+        return int(re.compile("._(\\d+)\\.pth").match(_f).group(1))
+
+    def time_key(_f):
+        return os.path.getmtime(os.path.join(path_to_models, _f))
+
+    sort_key = time_key if sort_by_time else name_key
+
+    def x_sorted(_x):
+        return sorted(
+            [f for f in ckpts_files if f.startswith(_x) and not f.endswith("_0.pth")],
+            key=sort_key,
+        )
+
+    to_del = [
+        os.path.join(path_to_models, fn)
+        for fn in (x_sorted("G")[:-n_ckpts_to_keep] + x_sorted("D")[:-n_ckpts_to_keep])
+    ]
+
+    def del_info(fn):
+        return logger.info(f".. Free up space by deleting ckpt {fn}")
+
+    def del_routine(x):
+        return [os.remove(x), del_info(x)]
+
+    [del_routine(fn) for fn in to_del]
+
+
+def get_hparams_from_dir(model_dir):
+    config_save_path = os.path.join(model_dir, "config.json")
+    with open(config_save_path, "r", encoding="utf-8") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    return hparams
+
+
+def get_hparams_from_file(config_path):
+    with open(config_path, "r", encoding="utf-8") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    return hparams
+
+
+def check_git_hash(model_dir):
+    source_dir = os.path.dirname(os.path.realpath(__file__))
+    if not os.path.exists(os.path.join(source_dir, ".git")):
+        logger.warn(
+            "{} is not a git repository, therefore hash value comparison will be ignored.".format(
+                source_dir
+            )
+        )
+        return
+
+    cur_hash = subprocess.getoutput("git rev-parse HEAD")
+
+    path = os.path.join(model_dir, "githash")
+    if os.path.exists(path):
+        saved_hash = open(path).read()
+        if saved_hash != cur_hash:
+            logger.warn(
+                "git hash values are different. {}(saved) != {}(current)".format(
+                    saved_hash[:8], cur_hash[:8]
+                )
+            )
+    else:
+        open(path, "w").write(cur_hash)
+
+
+def get_logger(model_dir, filename="train.log"):
+    global logger
+    logger = logging.getLogger(os.path.basename(model_dir))
+    logger.setLevel(logging.DEBUG)
+
+    formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir, exist_ok=True)
+    h = logging.FileHandler(os.path.join(model_dir, filename))
+    h.setLevel(logging.DEBUG)
+    h.setFormatter(formatter)
+    logger.addHandler(h)
+    return logger
+
+
+class HParams:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            if type(v) == dict:
+                v = HParams(**v)
+            self[k] = v
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def items(self):
+        return self.__dict__.items()
+
+    def values(self):
+        return self.__dict__.values()
+
+    def __len__(self):
+        return len(self.__dict__)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+    def __setitem__(self, key, value):
+        return setattr(self, key, value)
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __repr__(self):
+        return self.__dict__.__repr__()

src/vietnamese/__init__.py

@@ -0,0 +1,6 @@
+"""
+Vietnamese language support package
+"""
+
+from .phonemizer import text_to_phonemes, VIPHONEME_AVAILABLE, get_all_phonemes
+from .text_processor import process_vietnamese_text

src/vietnamese/phonemizer.py

@@ -0,0 +1,484 @@
+import atexit
+import contextlib
+import importlib.util
+import io
+import os
+import re
+import shutil
+import sys
+import tempfile
+import unicodedata
+from typing import List, Tuple
+from viphoneme import vi2IPA
+
+try:
+    import fcntl  # type: ignore
+except Exception:
+    fcntl = None
+
+VIPHONEME_AVAILABLE = True
+_VIPHONEME_WORKDIR = None
+_VINORM_ISOLATED_PARENT = None
+
+
+def _get_viphoneme_workdir() -> str:
+    global _VIPHONEME_WORKDIR
+    if _VIPHONEME_WORKDIR is None:
+        _VIPHONEME_WORKDIR = tempfile.mkdtemp(prefix="viphoneme_")
+        atexit.register(shutil.rmtree, _VIPHONEME_WORKDIR, ignore_errors=True)
+    return _VIPHONEME_WORKDIR
+
+
+def _ensure_vinorm_isolated() -> None:
+    global _VINORM_ISOLATED_PARENT
+    if os.environ.get("VIPHONEME_ISOLATE_VINORM", "1") not in {"1", "true", "True", "YES", "yes"}:
+        return
+    if _VINORM_ISOLATED_PARENT is not None:
+        return
+
+    spec = importlib.util.find_spec("vinorm")
+    if spec is None or spec.origin is None:
+        return
+
+    src_dir = os.path.dirname(spec.origin)
+    if not os.path.isfile(os.path.join(src_dir, "__init__.py")):
+        return
+
+    parent = tempfile.mkdtemp(prefix="vinorm_")
+    dst_dir = os.path.join(parent, "vinorm")
+    os.makedirs(dst_dir, exist_ok=True)
+
+    shutil.copy2(os.path.join(src_dir, "__init__.py"), os.path.join(dst_dir, "__init__.py"))
+
+    for name in os.listdir(src_dir):
+        if name in {"__init__.py", "__pycache__", "input.txt", "output.txt"}:
+            continue
+        src = os.path.join(src_dir, name)
+        dst = os.path.join(dst_dir, name)
+        if os.path.exists(dst):
+            continue
+        try:
+            os.symlink(src, dst)
+        except Exception:
+            if os.path.isdir(src):
+                shutil.copytree(src, dst)
+            elif os.path.isfile(src):
+                shutil.copy2(src, dst)
+
+    if parent not in sys.path:
+        sys.path.insert(0, parent)
+    if "vinorm" in sys.modules:
+        del sys.modules["vinorm"]
+
+    _VINORM_ISOLATED_PARENT = parent
+    atexit.register(shutil.rmtree, parent, ignore_errors=True)
+
+
+@contextlib.contextmanager
+def _redirect_fds_to_devnull():
+    devnull_fd = os.open(os.devnull, os.O_WRONLY)
+    saved_stdout_fd = os.dup(1)
+    saved_stderr_fd = os.dup(2)
+    try:
+        os.dup2(devnull_fd, 1)
+        os.dup2(devnull_fd, 2)
+        yield
+    finally:
+        os.dup2(saved_stdout_fd, 1)
+        os.dup2(saved_stderr_fd, 2)
+        os.close(saved_stdout_fd)
+        os.close(saved_stderr_fd)
+        os.close(devnull_fd)
+
+
+@contextlib.contextmanager
+def _viphoneme_global_lock():
+    lock_path = os.environ.get("VIPHONEME_LOCK_PATH", "/tmp/viphoneme.lock")
+    use_lock = os.environ.get("VIPHONEME_USE_LOCK")
+    if use_lock is None:
+        use_lock = "0" if os.environ.get("VIPHONEME_ISOLATE_VINORM", "1") in {"1", "true", "True", "YES", "yes"} else "1"
+    if use_lock not in {"1", "true", "True", "YES", "yes"}:
+        yield
+        return
+    if fcntl is None:
+        yield
+        return
+    fd = os.open(lock_path, os.O_CREAT | os.O_RDWR, 0o666)
+    try:
+        fcntl.flock(fd, fcntl.LOCK_EX)
+        yield
+    finally:
+        try:
+            fcntl.flock(fd, fcntl.LOCK_UN)
+        finally:
+            os.close(fd)
+
+
+# Vietnamese tone diacritics to tone number mapping
+TONE_MARKS = {
+    '\u0300': 2,  # ̀  huyền (falling)
+    '\u0301': 1,  # ́  sắc (rising) 
+    '\u0303': 3,  # ̃  ngã (broken)
+    '\u0309': 4,  # ̉  hỏi (dipping)
+    '\u0323': 5,  # ̣  nặng (heavy/glottalized)
+}
+
+# Default tone (no diacritic) = 0 (ngang/level)
+
+# Vietnamese orthography to IPA mapping
+VI_TO_IPA = {
+    # Trigraphs (check first)
+    'ngh': 'ŋ',
+    
+    # Digraphs
+    'ng': 'ŋ',
+    'nh': 'ɲ',
+    'ch': 'c',      # Vietnamese ch = palatal stop
+    'tr': 'ʈ',      # Retroflex
+    'th': 'tʰ',     # Aspirated
+    'ph': 'f',
+    'kh': 'x',      # Voiceless velar fricative
+    'gh': 'ɣ',
+    'gi': 'z',
+    'qu': 'kw',
+    
+    # Special consonants
+    'đ': 'ɗ',       # Implosive d
+    
+    # Simple consonants
+    'b': 'ɓ',       # Implosive b (can also be plain b)
+    'c': 'k',
+    'd': 'z',       # Northern: z, Southern: j
+    'g': 'ɣ',
+    'h': 'h',
+    'k': 'k',
+    'l': 'l',
+    'm': 'm',
+    'n': 'n',
+    'p': 'p',
+    'r': 'ʐ',       # Retroflex (varies by dialect)
+    's': 's',
+    't': 't',
+    'v': 'v',
+    'x': 's',       # Vietnamese x = s
+    
+    # Vowels
+    'a': 'aː',
+    'ă': 'a',       # Short a
+    'â': 'ə',       # Schwa
+    'e': 'ɛ',
+    'ê': 'e',
+    'i': 'i',
+    'y': 'i',       # Same as i
+    'o': 'ɔ',
+    'ô': 'o',
+    'ơ': 'əː',      # Long schwa
+    'u': 'u',
+    'ư': 'ɯ',       # Unrounded u
+    
+    # Diphthongs (handled separately)
+}
+
+# Final consonants (codas)
+FINAL_CONSONANTS = {
+    'c': 'k',
+    'ch': 'c',
+    'm': 'm',
+    'n': 'n',
+    'ng': 'ŋ',
+    'nh': 'ɲ',
+    'p': 'p',
+    't': 't',
+}
+
+# Punctuation to keep
+PUNCTUATION = set(',.!?;:\'"--—…()[]{}')
+
+# Punctuation that creates pauses (SP = short pause)
+PAUSE_PUNCTUATION = {',', ';', ':'}
+STOP_PUNCTUATION = {'.', '!', '?', '…'}
+
+def extract_tone(char: str) -> Tuple[str, int]:
+    """
+    Extract tone from a Vietnamese character.
+    Returns (base_char, tone_number)
+    """
+    # Decompose to separate base and combining marks
+    decomposed = unicodedata.normalize('NFD', char)
+    base = ''
+    tone = 0
+    
+    for c in decomposed:
+        if c in TONE_MARKS:
+            tone = TONE_MARKS[c]
+        elif not unicodedata.combining(c):
+            base += c
+    
+    return base, tone
+
+
+def syllable_to_ipa(syllable: str) -> Tuple[List[str], int]:
+    """
+    Convert a Vietnamese syllable to IPA phonemes with tone.
+    Returns (phonemes, tone)
+    """
+    syllable = syllable.lower()
+    phonemes = []
+    tone = 0
+    
+    # Extract tone from vowels
+    processed = ''
+    for char in syllable:
+        base, char_tone = extract_tone(char)
+        if char_tone > 0:
+            tone = char_tone
+        processed += base
+    
+    syllable = processed
+    i = 0
+    
+    while i < len(syllable):
+        matched = False
+        
+        # Try trigraphs
+        if i + 2 < len(syllable):
+            tri = syllable[i:i+3]
+            if tri in VI_TO_IPA:
+                phonemes.append(VI_TO_IPA[tri])
+                i += 3
+                matched = True
+        
+        # Try digraphs
+        if not matched and i + 1 < len(syllable):
+            di = syllable[i:i+2]
+            if di in VI_TO_IPA:
+                phonemes.append(VI_TO_IPA[di])
+                i += 2
+                matched = True
+        
+        # Single character
+        if not matched:
+            char = syllable[i]
+            if char in VI_TO_IPA:
+                phonemes.append(VI_TO_IPA[char])
+            elif char.isalpha():
+                phonemes.append(char)  # Keep as-is if not mapped
+            i += 1
+    
+    return phonemes, tone
+
+
+def text_to_phonemes_viphoneme(text: str) -> Tuple[List[str], List[int], List[int]]:
+    """
+    Convert text to phonemes using viphoneme library.
+    Returns (phones, tones, word2ph)
+    
+    viphoneme output format:
+    - Syllables separated by space
+    - Compound words joined by underscore: hom1_năj1
+    - Tone number (1-6) at end of each syllable
+    - Punctuation as separate tokens
+    """
+    import warnings
+    
+    # Call viphoneme (ICU warnings will appear but won't affect results)
+    # Note: viphoneme may not work on Windows due to platform-specific binaries
+    try:
+        _ensure_vinorm_isolated()
+        workdir = _get_viphoneme_workdir()
+        with _viphoneme_global_lock():
+            cwd = os.getcwd()
+            os.chdir(workdir)
+            try:
+                with warnings.catch_warnings():
+                    warnings.simplefilter("ignore")
+                    with _redirect_fds_to_devnull():
+                        ipa_text = vi2IPA(text)
+            finally:
+                os.chdir(cwd)
+    except Exception:
+        # Fallback to char-based on error (e.g., Windows compatibility issues)
+        return text_to_phonemes_charbased(text)
+    
+    # Check if viphoneme returned empty or invalid result
+    if not ipa_text or ipa_text.strip() in ['', '.', '..', '...']:
+        return text_to_phonemes_charbased(text)
+    
+    phones = []
+    tones = []
+    word2ph = []
+    
+    # viphoneme tone mapping: 1=ngang, 2=huyền, 3=ngã, 4=hỏi, 5=sắc, 6=nặng
+    # Our internal: 0=ngang, 1=sắc, 2=huyền, 3=ngã, 4=hỏi, 5=nặng
+    VIPHONEME_TONE_MAP = {1: 0, 2: 2, 3: 3, 4: 4, 5: 1, 6: 5}
+    
+    # Characters to skip (combining marks, ties)
+    SKIP_CHARS = {'\u0306', '\u0361', '\u032f', '\u0330', '\u0329'}  # breve, tie, etc.
+    
+    # Split by space
+    tokens = ipa_text.strip().split()
+    
+    for token in tokens:
+        # Handle punctuation-only tokens
+        if all(c in PUNCTUATION or c == '.' for c in token):
+            for c in token:
+                if c in PUNCTUATION:
+                    phones.append(c)
+                    tones.append(0)
+                    word2ph.append(1)
+            continue
+        
+        # Split compound words by underscore
+        syllables = token.split('_')
+        
+        for syllable in syllables:
+            if not syllable:
+                continue
+                
+            syllable_phones = []
+            syllable_tone = 0
+            i = 0
+            
+            while i < len(syllable):
+                char = syllable[i]
+                
+                # Tone number at end
+                if char.isdigit():
+                    syllable_tone = VIPHONEME_TONE_MAP.get(int(char), 0)
+                    i += 1
+                    continue
+                
+                # Skip combining marks (they modify previous char, already handled)
+                if unicodedata.combining(char):
+                    i += 1
+                    continue
+                
+                # Skip modifier letters like ʷ ʰ (append to previous if exists)
+                if char in {'ʷ', 'ʰ', 'ː'}:
+                    if syllable_phones:
+                        syllable_phones[-1] = syllable_phones[-1] + char
+                    i += 1
+                    continue
+                
+                # Skip tie bars and other special marks
+                if char in {'\u0361', '\u035c', '\u0361'}:  # tie bars
+                    i += 1
+                    continue
+                
+                # Punctuation within syllable
+                if char in PUNCTUATION:
+                    i += 1
+                    continue
+                
+                # Regular phoneme character
+                syllable_phones.append(char)
+                i += 1
+            
+            if syllable_phones:
+                phones.extend(syllable_phones)
+                tones.extend([syllable_tone] * len(syllable_phones))
+                word2ph.append(len(syllable_phones))
+    
+    return phones, tones, word2ph
+
+
+def text_to_phonemes_charbased(text: str) -> Tuple[List[str], List[int], List[int]]:
+    """
+    Convert text to phonemes using character-based mapping.
+    Returns (phones, tones, word2ph)
+    """
+    phones = []
+    tones = []
+    word2ph = []
+    
+    words = text.split()
+    
+    for word in words:
+        # Check for punctuation at end
+        trailing_punct = []
+        while word and word[-1] in PUNCTUATION:
+            trailing_punct.insert(0, word[-1])
+            word = word[:-1]
+        
+        # Check for punctuation at start
+        leading_punct = []
+        while word and word[0] in PUNCTUATION:
+            leading_punct.append(word[0])
+            word = word[1:]
+        
+        # Add leading punctuation
+        for p in leading_punct:
+            phones.append(p)
+            tones.append(0)
+            word2ph.append(1)
+        
+        # Process word syllables (Vietnamese words can be multi-syllable)
+        if word:
+            word_phones, tone = syllable_to_ipa(word)
+            if word_phones:
+                phones.extend(word_phones)
+                tones.extend([tone] * len(word_phones))
+                word2ph.append(len(word_phones))
+        
+        # Add trailing punctuation
+        for p in trailing_punct:
+            phones.append(p)
+            tones.append(0)
+            word2ph.append(1)
+    
+    return phones, tones, word2ph
+
+
+def text_to_phonemes(text: str, use_viphoneme: bool = True) -> Tuple[List[str], List[int], List[int]]:
+    """
+    Main function to convert Vietnamese text to phonemes.
+    
+    Args:
+        text: Vietnamese text
+        use_viphoneme: Whether to use viphoneme library (if available)
+        
+    Returns:
+        phones: List of IPA phonemes
+        tones: List of tone numbers (0-5)
+        word2ph: List of phone counts per word
+    """
+    if use_viphoneme and VIPHONEME_AVAILABLE:
+        phones, tones, word2ph = text_to_phonemes_viphoneme(text)
+    else:
+        phones, tones, word2ph = text_to_phonemes_charbased(text)
+    
+    # Add boundary tokens
+    phones = ["_"] + phones + ["_"]
+    tones = [0] + tones + [0]
+    word2ph = [1] + word2ph + [1]
+    
+    return phones, tones, word2ph
+
+
+def get_all_phonemes() -> List[str]:
+    """Get list of all possible phonemes for symbol table."""
+    phonemes = set()
+    
+    # From IPA mapping
+    for ipa in VI_TO_IPA.values():
+        if isinstance(ipa, str):
+            phonemes.add(ipa)
+            # Also add with length marker
+            if len(ipa) == 1:
+                phonemes.add(ipa + 'ː')
+    
+    # Common IPA symbols
+    phonemes.update([
+        # Consonants
+        'b', 'ɓ', 'c', 'd', 'ɗ', 'f', 'g', 'ɣ', 'h', 'j', 'k', 'l', 'm', 'n',
+        'ŋ', 'ɲ', 'p', 'r', 'ʐ', 's', 'ʂ', 't', 'tʰ', 'ʈ', 'v', 'w', 'x', 'z',
+        # Vowels
+        'a', 'aː', 'ə', 'əː', 'ɛ', 'e', 'i', 'ɪ', 'o', 'ɔ', 'u', 'ʊ', 'ɯ', 'ɤ',
+        # Special
+        '_', ' ',
+    ])
+    
+    # Punctuation
+    phonemes.update(PUNCTUATION)
+    
+    return sorted(list(phonemes))

src/vietnamese/text_processor.py

@@ -0,0 +1,428 @@
+#!/usr/bin/env python3
+"""
+Vietnamese Text Processor for TTS
+Handles normalization of numbers, dates, times, currencies, etc.
+"""
+
+import re
+import unicodedata
+
+
+# Vietnamese number words
+DIGITS = {
+    '0': 'không', '1': 'một', '2': 'hai', '3': 'ba', '4': 'bốn',
+    '5': 'năm', '6': 'sáu', '7': 'bảy', '8': 'tám', '9': 'chín'
+}
+
+TEENS = {
+    '10': 'mười', '11': 'mười một', '12': 'mười hai', '13': 'mười ba',
+    '14': 'mười bốn', '15': 'mười lăm', '16': 'mười sáu', '17': 'mười bảy',
+    '18': 'mười tám', '19': 'mười chín'
+}
+
+TENS = {
+    '2': 'hai mươi', '3': 'ba mươi', '4': 'bốn mươi', '5': 'năm mươi',
+    '6': 'sáu mươi', '7': 'bảy mươi', '8': 'tám mươi', '9': 'chín mươi'
+}
+
+
+def number_to_words(num_str):
+    """
+    Convert a number string to Vietnamese words.
+    Handles numbers from 0 to billions.
+    """
+    # Remove leading zeros but keep at least one digit
+    num_str = num_str.lstrip('0') or '0'
+    
+    # Handle negative numbers
+    if num_str.startswith('-'):
+        return 'âm ' + number_to_words(num_str[1:])
+    
+    # Convert to integer for processing
+    try:
+        num = int(num_str)
+    except ValueError:
+        return num_str
+    
+    if num == 0:
+        return 'không'
+    
+    if num < 10:
+        return DIGITS[str(num)]
+    
+    if num < 20:
+        return TEENS[str(num)]
+    
+    if num < 100:
+        tens = num // 10
+        units = num % 10
+        if units == 0:
+            return TENS[str(tens)]
+        elif units == 1:
+            return TENS[str(tens)] + ' mốt'
+        elif units == 4:
+            return TENS[str(tens)] + ' tư'
+        elif units == 5:
+            return TENS[str(tens)] + ' lăm'
+        else:
+            return TENS[str(tens)] + ' ' + DIGITS[str(units)]
+    
+    if num < 1000:
+        hundreds = num // 100
+        remainder = num % 100
+        result = DIGITS[str(hundreds)] + ' trăm'
+        if remainder == 0:
+            return result
+        elif remainder < 10:
+            return result + ' lẻ ' + DIGITS[str(remainder)]
+        else:
+            return result + ' ' + number_to_words(str(remainder))
+    
+    if num < 1000000:
+        thousands = num // 1000
+        remainder = num % 1000
+        result = number_to_words(str(thousands)) + ' nghìn'
+        if remainder == 0:
+            return result
+        elif remainder < 100:
+            return result + ' không trăm ' + number_to_words(str(remainder))
+        else:
+            return result + ' ' + number_to_words(str(remainder))
+    
+    if num < 1000000000:
+        millions = num // 1000000
+        remainder = num % 1000000
+        result = number_to_words(str(millions)) + ' triệu'
+        if remainder == 0:
+            return result
+        else:
+            return result + ' ' + number_to_words(str(remainder))
+    
+    if num < 1000000000000:
+        billions = num // 1000000000
+        remainder = num % 1000000000
+        result = number_to_words(str(billions)) + ' tỷ'
+        if remainder == 0:
+            return result
+        else:
+            return result + ' ' + number_to_words(str(remainder))
+    
+    # For very large numbers, read digit by digit
+    return ' '.join(DIGITS.get(d, d) for d in num_str)
+
+
+def convert_decimal(text):
+    """Convert decimal numbers: 3.14 -> ba phẩy mười bốn"""
+    def replace_decimal(match):
+        integer_part = match.group(1)
+        decimal_part = match.group(2)
+        
+        integer_words = number_to_words(integer_part)
+        
+        # Read decimal part as a number
+        decimal_words = number_to_words(decimal_part.lstrip('0') or '0')
+        
+        return f"{integer_words} phẩy {decimal_words}"
+    
+    # Match decimal numbers: X.Y where Y is 1-2 digits, followed by space or end
+    # Avoid matching large numbers like 100.000 (thousand separator)
+    text = re.sub(r'(\d+)\.(\d{1,2})(?=\s|$|[^\d])', replace_decimal, text)
+    return text
+
+
+def convert_percentage(text):
+    """Convert percentages: 50% -> năm mươi phần trăm"""
+    def replace_percent(match):
+        num = match.group(1)
+        return number_to_words(num) + ' phần trăm'
+    
+    text = re.sub(r'(\d+(?:[.,]\d+)?)\s*%', replace_percent, text)
+    return text
+
+
+def convert_currency(text):
+    """Convert currency amounts"""
+    # Vietnamese Dong - be specific to avoid matching "đ" in other words like "độ"
+    def replace_vnd(match):
+        num = match.group(1).replace('.', '').replace(',', '')
+        return number_to_words(num) + ' đồng'
+    
+    # Only match currency patterns: number followed by currency symbol at word boundary
+    text = re.sub(r'(\d+(?:[.,]\d+)*)\s*(?:đồng|VND|vnđ)\b', replace_vnd, text, flags=re.IGNORECASE)
+    text = re.sub(r'(\d+(?:[.,]\d+)*)đ(?![a-zà-ỹ])', replace_vnd, text, flags=re.IGNORECASE)
+    
+    # USD
+    def replace_usd(match):
+        num = match.group(1).replace('.', '').replace(',', '')
+        return number_to_words(num) + ' đô la'
+    
+    text = re.sub(r'\$\s*(\d+(?:[.,]\d+)*)', replace_usd, text)
+    text = re.sub(r'(\d+(?:[.,]\d+)*)\s*(?:USD|\$)', replace_usd, text, flags=re.IGNORECASE)
+    
+    return text
+
+
+def convert_time(text):
+    """Convert time expressions: 2 giờ 20 phút -> hai giờ hai mươi phút"""
+    def replace_time(match):
+        hour = match.group(1)
+        minute = match.group(2) if match.group(2) else None
+        second = match.group(3) if len(match.groups()) > 2 and match.group(3) else None
+        
+        result = number_to_words(hour) + ' giờ'
+        if minute:
+            result += ' ' + number_to_words(minute) + ' phút'
+        if second:
+            result += ' ' + number_to_words(second) + ' giây'
+        return result
+    
+    # HH:MM:SS or HH:MM
+    text = re.sub(r'(\d{1,2}):(\d{2})(?::(\d{2}))?', replace_time, text)
+    
+    # X giờ Y phút
+    def replace_time_vn(match):
+        hour = match.group(1)
+        minute = match.group(2)
+        return number_to_words(hour) + ' giờ ' + number_to_words(minute) + ' phút'
+    
+    text = re.sub(r'(\d+)\s*giờ\s*(\d+)\s*phút', replace_time_vn, text)
+    
+    # X giờ (without minute)
+    def replace_hour(match):
+        hour = match.group(1)
+        return number_to_words(hour) + ' giờ'
+    
+    text = re.sub(r'(\d+)\s*giờ(?!\s*\d)', replace_hour, text)
+    
+    return text
+
+
+def convert_date(text):
+    """Convert date expressions"""
+    # DD/MM/YYYY or DD-MM-YYYY
+    def replace_date_full(match):
+        day = match.group(1)
+        month = match.group(2)
+        year = match.group(3)
+        return f"ngày {number_to_words(day)} tháng {number_to_words(month)} năm {number_to_words(year)}"
+    
+    # First, replace "Sinh ngày DD/MM/YYYY" pattern to avoid double "ngày"
+    text = re.sub(r'(Sinh|sinh)\s+ngày\s+(\d{1,2})[/-](\d{1,2})[/-](\d{4})', 
+                  lambda m: f"{m.group(1)} ngày {number_to_words(m.group(2))} tháng {number_to_words(m.group(3))} năm {number_to_words(m.group(4))}", text)
+    
+    text = re.sub(r'(\d{1,2})[/-](\d{1,2})[/-](\d{4})', replace_date_full, text)
+    
+    # X tháng Y
+    def replace_month_day(match):
+        day = match.group(1)
+        month = match.group(2)
+        return f"ngày {number_to_words(day)} tháng {number_to_words(month)}"
+    
+    text = re.sub(r'(\d+)\s*tháng\s*(\d+)', replace_month_day, text)
+    
+    # tháng X (month only)
+    def replace_month(match):
+        month = match.group(1)
+        return 'tháng ' + number_to_words(month)
+    
+    text = re.sub(r'tháng\s*(\d+)', replace_month, text)
+    
+    # ngày X
+    def replace_day(match):
+        day = match.group(1)
+        return 'ngày ' + number_to_words(day)
+    
+    text = re.sub(r'ngày\s*(\d+)', replace_day, text)
+    
+    return text
+
+
+def convert_year_range(text):
+    """Convert year ranges: 1873-1907 -> một nghìn tám trăm bảy mươi ba đến một nghìn chín trăm lẻ bảy"""
+    def replace_year_range(match):
+        year1 = match.group(1)
+        year2 = match.group(2)
+        return number_to_words(year1) + ' đến ' + number_to_words(year2)
+    
+    text = re.sub(r'(\d{4})\s*[-–—]\s*(\d{4})', replace_year_range, text)
+    return text
+
+
+def convert_ordinal(text):
+    """Convert ordinals: thứ 2 -> thứ hai"""
+    ordinal_map = {
+        '1': 'nhất', '2': 'hai', '3': 'ba', '4': 'tư', '5': 'năm',
+        '6': 'sáu', '7': 'bảy', '8': 'tám', '9': 'chín', '10': 'mười'
+    }
+    
+    def replace_ordinal(match):
+        prefix = match.group(1)
+        num = match.group(2)
+        if num in ordinal_map:
+            return prefix + ' ' + ordinal_map[num]
+        return prefix + ' ' + number_to_words(num)
+    
+    # thứ X, lần X, bước X, phần X
+    text = re.sub(r'(thứ|lần|bước|phần|chương|tập|số)\s*(\d+)', replace_ordinal, text, flags=re.IGNORECASE)
+    
+    return text
+
+
+def convert_standalone_numbers(text):
+    """Convert remaining standalone numbers to words"""
+    def replace_num(match):
+        num = match.group(0)
+        # Skip if it's part of a word or already processed
+        return number_to_words(num)
+    
+    # Match numbers not followed/preceded by letters
+    text = re.sub(r'\b\d+\b', replace_num, text)
+    return text
+
+
+def convert_phone_number(text):
+    """Read phone numbers digit by digit"""
+    def replace_phone(match):
+        phone = match.group(0)
+        digits = re.findall(r'\d', phone)
+        return ' '.join(DIGITS.get(d, d) for d in digits)
+    
+    # Vietnamese phone patterns
+    text = re.sub(r'0\d{9,10}', replace_phone, text)
+    text = re.sub(r'\+84\d{9,10}', replace_phone, text)
+    
+    return text
+
+
+def normalize_unicode(text):
+    """Normalize Unicode to NFC form"""
+    return unicodedata.normalize('NFC', text)
+
+
+def clean_whitespace(text):
+    """Clean up extra whitespace"""
+    text = re.sub(r'\s+', ' ', text)
+    return text.strip()
+
+
+def remove_special_chars(text):
+    """Remove or replace special characters that can't be spoken"""
+    # Keep Vietnamese diacritics and common punctuation
+    # Remove emojis and special symbols
+    
+    # Replace common symbols with words
+    text = text.replace('&', ' và ')
+    text = text.replace('@', ' a còng ')
+    text = text.replace('#', ' thăng ')
+    text = text.replace('*', '')
+    text = text.replace('_', ' ')
+    text = text.replace('~', '')
+    text = text.replace('`', '')
+    text = text.replace('^', '')
+    
+    # Remove URLs
+    text = re.sub(r'https?://\S+', '', text)
+    text = re.sub(r'www\.\S+', '', text)
+    
+    # Remove email addresses
+    text = re.sub(r'\S+@\S+\.\S+', '', text)
+    
+    return text
+
+
+def normalize_punctuation(text):
+    """Normalize punctuation marks"""
+    # Normalize quotes
+    text = re.sub(r'[""„‟]', '"', text)
+    text = re.sub(r"[''‚‛]", "'", text)
+    
+    # Normalize dashes
+    text = re.sub(r'[–—−]', '-', text)
+    
+    # Normalize ellipsis
+    text = re.sub(r'\.{3,}', '...', text)
+    text = text.replace('…', '...')
+    
+    # Remove multiple punctuation
+    text = re.sub(r'([!?.]){2,}', r'\1', text)
+    
+    return text
+
+
+def process_vietnamese_text(text):
+    """
+    Main function to process Vietnamese text for TTS.
+    Applies all normalization steps in the correct order.
+    
+    Args:
+        text: Raw Vietnamese text
+        
+    Returns:
+        Normalized text suitable for TTS
+    """
+    # Step 1: Normalize Unicode
+    text = normalize_unicode(text)
+    
+    # Step 2: Remove special characters
+    text = remove_special_chars(text)
+    
+    # Step 3: Normalize punctuation
+    text = normalize_punctuation(text)
+    
+    # Step 4: Convert year ranges (before other number conversions)
+    text = convert_year_range(text)
+    
+    # Step 5: Convert dates
+    text = convert_date(text)
+    
+    # Step 6: Convert times
+    text = convert_time(text)
+    
+    # Step 7: Convert ordinals
+    text = convert_ordinal(text)
+    
+    # Step 8: Convert currency
+    text = convert_currency(text)
+    
+    # Step 9: Convert percentages
+    text = convert_percentage(text)
+    
+    # Step 10: Convert phone numbers
+    text = convert_phone_number(text)
+    
+    # Step 11: Convert decimals (before standalone numbers, after currency)
+    text = convert_decimal(text)
+    
+    # Step 12: Convert remaining standalone numbers
+    text = convert_standalone_numbers(text)
+    
+    # Step 13: Clean whitespace
+    text = clean_whitespace(text)
+    
+    return text
+
+
+if __name__ == "__main__":
+    # Test cases
+    test_cases = [
+        "Lúc khoảng 2 giờ 20 phút sáng ngày thứ Bảy hay 8 tháng 11",
+        "Alfred Jarry 1873-1907 hợp những nhà văn",
+        "ông Derringer 44 ly, dí sát đầu tổng thống",
+        "Giá sản phẩm là 100.000đ",
+        "Tỷ lệ thành công đạt 85%",
+        "Họp lúc 14:30",
+        "Sinh ngày 15/08/1990",
+        "Chương 3: Hành trình mới",
+        "Số điện thoại: 0912345678",
+        "Nhiệt độ 25.5 độ C",
+        "Công ty XYZ có 1500 nhân viên",
+    ]
+    
+    print("=" * 60)
+    print("Vietnamese Text Processor Test")
+    print("=" * 60)
+    
+    for text in test_cases:
+        processed = process_vietnamese_text(text)
+        print(f"\nOriginal: {text}")
+        print(f"Processed: {processed}")