app.py

import torch
import torchaudio
import gradio as gr
from fastapi import FastAPI
from speechbrain.pretrained import Tacotron2, HIFIGAN

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
SR = 22050

tacotron2 = Tacotron2.from_hparams(
    source="speechbrain/tts-tacotron2-ljspeech",
    savedir="models/tacotron2",
    run_opts={"device": DEVICE},
)

hifigan = HIFIGAN.from_hparams(
    source="speechbrain/tts-hifigan-ljspeech",
    savedir="models/hifigan",
    run_opts={"device": DEVICE},
)

@torch.inference_mode()
def synth(text):
    text = (text or "").strip()
    if not text:
        return None

    seq, seq_len = tacotron2.text_to_seq(text)
    seq = [int(x) for x in seq]

    # Tacotron2 encoder conv needs enough timesteps
    min_tokens = 5
    pad_id = 0
    if len(seq) < min_tokens:
        seq = seq + [pad_id] * (min_tokens - len(seq))
        seq_len = len(seq)

    seq = torch.tensor(seq, dtype=torch.long, device=DEVICE).unsqueeze(0)
    seq_len = torch.tensor([seq_len], device=DEVICE)

    mel, _, _ = tacotron2.infer(seq, seq_len)

    # Optional: still keep mel padding for vocoder safety
    if mel.shape[-1] < 5:
        mel = F.pad(mel, (0, 5 - mel.shape[-1]), mode="replicate")

    wav = hifigan.decode_batch(mel)
    if wav.dim() == 3:
        wav = wav.squeeze(1)
    wav = wav[0].cpu()

    out_path = "out.wav"
    torchaudio.save(out_path, wav.unsqueeze(0), SR)
    return out_path


io = gr.Interface(
    fn=synth,
    inputs=gr.Textbox(label="Text", lines=3),
    outputs=gr.Audio(type="filepath", label="Output"),
    title="Tacotron 2 + HiFi-GAN",
    allow_flagging="never",
    api_name=False,
)

app = FastAPI()
app = gr.mount_gradio_app(app, io, path="/")