app.py

import gradio as gr
import numpy as np
import matplotlib.pyplot as plt
import torch
import yaml
import json
import pyloudnorm as pyln
from hydra.utils import instantiate
from soxr import resample
from functools import partial

from modules.utils import chain_functions, vec2statedict, get_chunks
from modules.fx import clip_delay_eq_Q
from plot_utils import get_log_mags_from_eq


title_md = "# Vocal Effects Generator"
description_md = """
This is a demo of the paper [DiffVox: A Differentiable Model for Capturing and Analysing Professional Effects Distributions](https://arxiv.org/abs/2504.14735), accepted at DAFx 2025.
In this demo, you can upload a raw vocal audio file (in mono) and apply random effects to make it sound better!

The effects consist of series of EQ, compressor, delay, and reverb.
The generator is a PCA model derived from 365 vocal effects presets fitted with the same effects chain.
This interface allows you to control the principal components (PCs) of the generator, randomise them, and render the audio.

To give you some idea, we emperically found that the first PC controls the amount of reverb and the second PC controls the amount of brightness.
Note that adding these PCs together does not necessarily mean that their effects are additive in the final audio.
We found sometimes the effects of least important PCs are more perceptible.
Try to play around with the sliders and buttons and see what you can come up with!

Currently only PCs are tweakable, but in the future we will add more controls and visualisation tools.
For example:
- Directly controlling the parameters of the effects
- Visualising the PCA space
- Visualising the frequency responses/dynamic curves of the effects
"""

SLIDER_MAX = 3
SLIDER_MIN = -3
NUMBER_OF_PCS = 4
TEMPERATURE = 0.7
CONFIG_PATH = "presets/rt_config.yaml"
PCA_PARAM_FILE = "presets/internal/gaussian.npz"
INFO_PATH = "presets/internal/info.json"
MASK_PATH = "presets/internal/feature_mask.npy"


with open(CONFIG_PATH) as fp:
    fx_config = yaml.safe_load(fp)["model"]

# Global effect
fx = instantiate(fx_config)
fx.eval()

pca_params = np.load(PCA_PARAM_FILE)
mean = pca_params["mean"]
cov = pca_params["cov"]
eigvals, eigvecs = np.linalg.eigh(cov)
eigvals = np.flip(eigvals, axis=0)
eigvecs = np.flip(eigvecs, axis=1)
U = eigvecs * np.sqrt(eigvals)
U = torch.from_numpy(U).float()
mean = torch.from_numpy(mean).float()
feature_mask = torch.from_numpy(np.load(MASK_PATH))
# Global latent variable
z = torch.zeros_like(mean)

with open(INFO_PATH) as f:
    info = json.load(f)

param_keys = info["params_keys"]
original_shapes = list(
    map(lambda lst: lst if len(lst) else [1], info["params_original_shapes"])
)

*vec2dict_args, _ = get_chunks(param_keys, original_shapes)
vec2dict_args = [param_keys, original_shapes] + vec2dict_args
vec2dict = partial(
    vec2statedict,
    **dict(
        zip(
            [
                "keys",
                "original_shapes",
                "selected_chunks",
                "position",
                "U_matrix_shape",
            ],
            vec2dict_args,
        )
    ),
)
fx.load_state_dict(vec2dict(mean), strict=False)


meter = pyln.Meter(44100)


@torch.no_grad()
def z2fx():
    # close all figures to avoid too many open figures
    plt.close("all")
    x = U @ z + mean
    # print(z)
    fx.load_state_dict(vec2dict(x), strict=False)
    fx.apply(partial(clip_delay_eq_Q, Q=0.707))
    return


@torch.no_grad()
def fx2z():
    state_dict = fx.state_dict()
    flattened = torch.cat([state_dict[k].flatten() for k in param_keys])
    x = flattened[feature_mask]
    z.copy_(U.T @ (x - mean))
    return


@torch.no_grad()
def inference(audio):
    sr, y = audio
    if sr != 44100:
        y = resample(y, sr, 44100)
    if y.dtype.kind != "f":
        y = y / 32768.0

    if y.ndim == 1:
        y = y[:, None]
    loudness = meter.integrated_loudness(y)
    y = pyln.normalize.loudness(y, loudness, -18.0)

    y = torch.from_numpy(y).float().T.unsqueeze(0)
    if y.shape[1] != 1:
        y = y.mean(dim=1, keepdim=True)

    rendered = fx(y).squeeze(0).T.numpy()
    if np.max(np.abs(rendered)) > 1:
        rendered = rendered / np.max(np.abs(rendered))
    return (44100, (rendered * 32768).astype(np.int16))


def get_important_pcs(n=10, **kwargs):
    sliders = [
        gr.Slider(minimum=SLIDER_MIN, maximum=SLIDER_MAX, label=f"PC {i}", **kwargs)
        for i in range(1, n + 1)
    ]
    return sliders


def model2json():
    fx_names = ["PK1", "PK2", "LS", "HS", "LP", "HP", "DRC"]
    results = {k: v.toJSON() for k, v in zip(fx_names, fx)} | {
        "Panner": fx[7].pan.toJSON()
    }
    spatial_fx = {
        "DLY": fx[7].effects[0].toJSON() | {"LP": fx[7].effects[0].eq.toJSON()},
        "FDN": fx[7].effects[1].toJSON()
        | {
            "Tone correction PEQ": {
                k: v.toJSON() for k, v in zip(fx_names[:4], fx[7].effects[1].eq)
            }
        },
        "Cross Send (dB)": fx[7].params.sends_0.log10().mul(20).item(),
    }
    replace_neg_inf = lambda d: (
        {k: (replace_neg_inf(v) if v != -np.inf else -1e500) for k, v in d.items()}
        if isinstance(d, dict)
        else d
    )
    return {
        "Direct": results,
        "Sends": spatial_fx,
    }


@torch.no_grad()
def plot_eq():
    fig, ax = plt.subplots(figsize=(6, 4), constrained_layout=True)
    w, eq_log_mags = get_log_mags_from_eq(fx[:6])
    ax.plot(w, sum(eq_log_mags), color="black", linestyle="-")
    for i, eq_log_mag in enumerate(eq_log_mags):
        ax.plot(w, eq_log_mag, "k-", alpha=0.3)
        ax.fill_between(w, eq_log_mag, 0, facecolor="gray", edgecolor="none", alpha=0.1)
    ax.set_xlabel("Frequency (Hz)")
    ax.set_ylabel("Magnitude (dB)")
    ax.set_xlim(20, 20000)
    ax.set_ylim(-40, 20)
    ax.set_xscale("log")
    ax.grid()
    return fig


@torch.no_grad()
def plot_comp():
    fig, ax = plt.subplots(figsize=(6, 5), constrained_layout=True)
    comp = fx[6]
    cmp_th = comp.params.cmp_th.item()
    exp_th = comp.params.exp_th.item()
    cmp_ratio = comp.params.cmp_ratio.item()
    exp_ratio = comp.params.exp_ratio.item()
    make_up = comp.params.make_up.item()
    # print(cmp_ratio, cmp_th, exp_ratio, exp_th, make_up)

    comp_in = np.linspace(-80, 0, 100)
    comp_curve = np.where(
        comp_in > cmp_th,
        comp_in - (comp_in - cmp_th) * (cmp_ratio - 1) / cmp_ratio,
        comp_in,
    )
    comp_out = (
        np.where(
            comp_curve < exp_th,
            comp_curve - (exp_th - comp_curve) / exp_ratio,
            comp_curve,
        )
        + make_up
    )
    ax.plot(comp_in, comp_out, c="black", linestyle="-")
    ax.plot(comp_in, comp_in, c="r", alpha=0.5)
    ax.set_xlabel("Input Level (dB)")
    ax.set_ylabel("Output Level (dB)")
    ax.set_xlim(-80, 0)
    ax.set_ylim(-80, 0)
    ax.grid()
    return fig


@torch.no_grad()
def plot_delay():
    fig, ax = plt.subplots(figsize=(6, 4), constrained_layout=True)
    delay = fx[7].effects[0]
    w, eq_log_mags = get_log_mags_from_eq([delay.eq])
    log_gain = delay.params.gain.log10().item() * 20
    d = delay.params.delay.item() / 1000
    log_mag = sum(eq_log_mags)
    ax.plot(w, log_mag + log_gain, color="black", linestyle="-")

    log_feedback = delay.params.feedback.log10().item() * 20
    for i in range(1, 10):
        feedback_log_mag = log_mag * (i + 1) + log_feedback * i + log_gain
        ax.plot(
            w,
            feedback_log_mag,
            c="black",
            alpha=max(0, (10 - i * d * 4) / 10),
            linestyle="-",
        )

    ax.set_xscale("log")
    ax.set_xlim(20, 20000)
    ax.set_ylim(-80, 0)
    ax.set_xlabel("Frequency (Hz)")
    ax.set_ylabel("Magnitude (dB)")
    ax.grid()
    return fig


@torch.no_grad()
def plot_reverb():
    fig, ax = plt.subplots(figsize=(6, 4), constrained_layout=True)
    fdn = fx[7].effects[1]
    w, eq_log_mags = get_log_mags_from_eq(fdn.eq)

    bc = fdn.params.c.norm() * fdn.params.b.norm()
    log_bc = torch.log10(bc).item() * 20
    eq_log_mags = [x + log_bc / len(eq_log_mags) for x in eq_log_mags]
    ax.plot(w, sum(eq_log_mags), color="black", linestyle="-")

    ax.set_xlabel("Frequency (Hz)")
    ax.set_ylabel("Magnitude (dB)")
    ax.set_xlim(20, 20000)
    ax.set_ylim(-40, 6)
    ax.set_xscale("log")
    ax.grid()
    return fig


@torch.no_grad()
def plot_t60():
    fig, ax = plt.subplots(figsize=(6, 4), constrained_layout=True)
    fdn = fx[7].effects[1]
    gamma = fdn.params.gamma.squeeze().numpy()
    delays = fdn.delays.numpy()
    w = np.linspace(0, 22050, gamma.size)
    t60 = -60 / (20 * np.log10(gamma + 1e-10) / np.min(delays)) / 44100
    ax.plot(w, t60, color="black", linestyle="-")
    ax.set_xlabel("Frequency (Hz)")
    ax.set_ylabel("T60 (s)")
    ax.set_xlim(20, 20000)
    ax.set_ylim(0, 9)
    ax.set_xscale("log")
    ax.grid()
    return fig


@torch.no_grad()
def upatePEQ(eq, attr_name, value):
    match type(getattr(eq.params, attr_name)):
        case torch.nn.Parameter:
            getattr(eq.params, attr_name).data.copy_(value)
        case _:
            setattr(eq.params, attr_name, torch.tensor(value))


with gr.Blocks() as demo:
    gr.Markdown(
        title_md,
        elem_id="title",
    )
    with gr.Row():
        gr.Markdown(
            description_md,
            elem_id="description",
        )
        gr.Image("diffvox_diagram.png", elem_id="diagram")

    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(
                type="numpy", sources="upload", label="Input Audio", loop=True
            )
            with gr.Row():
                random_button = gr.Button(
                    f"Randomise PCs",
                    elem_id="randomise-button",
                )
                reset_button = gr.Button(
                    "Reset",
                    elem_id="reset-button",
                )
                render_button = gr.Button(
                    "Run", elem_id="render-button", variant="primary"
                )
            # random_rest_checkbox = gr.Checkbox(
            #     label=f"Randomise PCs > {NUMBER_OF_PCS} (default to zeros)",
            #     value=False,
            #     elem_id="randomise-checkbox",
            # )
            # sliders = get_important_pcs(NUMBER_OF_PCS, value=0)
            with gr.Row():
                s1 = gr.Slider(
                    minimum=SLIDER_MIN,
                    maximum=SLIDER_MAX,
                    label="PC 1",
                    value=0,
                    interactive=True,
                )
                s2 = gr.Slider(
                    minimum=SLIDER_MIN,
                    maximum=SLIDER_MAX,
                    label="PC 2",
                    value=0,
                    interactive=True,
                )

            with gr.Row():
                s3 = gr.Slider(
                    minimum=SLIDER_MIN,
                    maximum=SLIDER_MAX,
                    label="PC 3",
                    value=0,
                    interactive=True,
                )
                s4 = gr.Slider(
                    minimum=SLIDER_MIN,
                    maximum=SLIDER_MAX,
                    label="PC 4",
                    value=0,
                    interactive=True,
                )

            sliders = [s1, s2, s3, s4]

            extra_pc_dropdown = gr.Dropdown(
                list(range(NUMBER_OF_PCS + 1, mean.numel() + 1)),
                label=f"PC > {NUMBER_OF_PCS}",
                info="Select which extra PC to adjust",
                interactive=True,
            )
            extra_slider = gr.Slider(
                minimum=SLIDER_MIN,
                maximum=SLIDER_MAX,
                label="Extra PC",
                value=0,
            )

        with gr.Column():
            audio_output = gr.Audio(
                type="numpy", label="Output Audio", interactive=False, loop=True
            )

            _ = gr.Markdown("## Parametric EQ")
            peq_plot = gr.Plot(
                plot_eq(), label="PEQ Frequency Response", elem_id="peq-plot"
            )
    with gr.Row():
        with gr.Column(min_width=160):
            _ = gr.Markdown("High Pass")
            hp = fx[5]
            hp_freq = gr.Slider(
                minimum=16,
                maximum=5300,
                value=hp.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            hp_q = gr.Slider(
                minimum=0.5,
                maximum=10,
                value=hp.params.Q.item(),
                interactive=True,
                label="Q",
            )

        with gr.Column(min_width=160):
            _ = gr.Markdown("Low Shelf")
            ls = fx[2]
            ls_freq = gr.Slider(
                minimum=30,
                maximum=200,
                value=ls.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            ls_gain = gr.Slider(
                minimum=-12,
                maximum=12,
                value=ls.params.gain.item(),
                interactive=True,
                label="Gain (dB)",
            )

        with gr.Column(min_width=160):
            _ = gr.Markdown("Peak filter 1")
            pk1 = fx[0]
            pk1_freq = gr.Slider(
                minimum=33,
                maximum=5400,
                value=pk1.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            pk1_gain = gr.Slider(
                minimum=-12,
                maximum=12,
                value=pk1.params.gain.item(),
                interactive=True,
                label="Gain (dB)",
            )
            pk1_q = gr.Slider(
                minimum=0.2,
                maximum=20,
                value=pk1.params.Q.item(),
                interactive=True,
                label="Q",
            )
        with gr.Column(min_width=160):
            _ = gr.Markdown("Peak filter 2")
            pk2 = fx[1]
            pk2_freq = gr.Slider(
                minimum=200,
                maximum=17500,
                value=pk2.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            pk2_gain = gr.Slider(
                minimum=-12,
                maximum=12,
                value=pk2.params.gain.item(),
                interactive=True,
                label="Gain (dB)",
            )
            pk2_q = gr.Slider(
                minimum=0.2,
                maximum=20,
                value=pk2.params.Q.item(),
                interactive=True,
                label="Q",
            )

        with gr.Column(min_width=160):
            _ = gr.Markdown("High Shelf")
            hs = fx[3]
            hs_freq = gr.Slider(
                minimum=750,
                maximum=8300,
                value=hs.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            hs_gain = gr.Slider(
                minimum=-12,
                maximum=12,
                value=hs.params.gain.item(),
                interactive=True,
                label="Gain (dB)",
            )
        with gr.Column(min_width=160):
            _ = gr.Markdown("Low Pass")
            lp = fx[4]
            lp_freq = gr.Slider(
                minimum=200,
                maximum=18000,
                value=lp.params.freq.item(),
                interactive=True,
                label="Frequency (Hz)",
            )
            lp_q = gr.Slider(
                minimum=0.5,
                maximum=10,
                value=lp.params.Q.item(),
                interactive=True,
                label="Q",
            )

    comp_plot = gr.Plot(plot_comp(), label="Compressor Curve", elem_id="comp-plot")
    delay_plot = gr.Plot(
        plot_delay(), label="Delay Frequency Response", elem_id="delay-plot"
    )
    reverb_plot = gr.Plot(
        plot_reverb(), label="Reverb Tone Correction PEQ", elem_id="reverb-plot"
    )
    t60_plot = gr.Plot(plot_t60(), label="Reverb T60", elem_id="t60-plot")

    with gr.Row():
        json_output = gr.JSON(
            model2json(), label="Effect Settings", max_height=800, open=True
        )

    update_pc = lambda i: z[:NUMBER_OF_PCS].tolist() + [z[i - 1].item()]
    update_pc_outputs = sliders + [extra_slider]

    for eq, s, attr_name in zip(
        [fx[0]] * 3
        + [fx[1]] * 3
        + [fx[2]] * 2
        + [fx[3]] * 2
        + [fx[4]] * 2
        + [fx[5]] * 2,
        [
            pk1_freq,
            pk1_gain,
            pk1_q,
            pk2_freq,
            pk2_gain,
            pk2_q,
            ls_freq,
            ls_gain,
            hs_freq,
            hs_gain,
            lp_freq,
            lp_q,
            hp_freq,
            hp_q,
        ],
        ["freq", "gain", "Q"] * 2 + ["freq", "gain"] * 2 + ["freq", "Q"] * 2,
    ):
        s.input(
            lambda *args, eq=eq, attr_name=attr_name: chain_functions(  # chain_functions(
                lambda args: (upatePEQ(eq, attr_name, args[0]), args[1]),
                lambda args: (fx2z(), args[1]),
                lambda args: args[1],
                lambda i: update_pc(i) + [model2json(), plot_eq()],
            )(
                args
            ),
            inputs=[s, extra_pc_dropdown],
            outputs=update_pc_outputs + [json_output, peq_plot],
        )

    render_button.click(
        # lambda *args: (
        #     lambda x: (
        #         x,
        #         model2json(),
        #     )
        # )(inference(*args)),
        inference,
        inputs=[
            audio_input,
        ],
        outputs=[
            audio_output,
        ],
    )

    update_fx = lambda: [
        pk1.params.freq.item(),
        pk1.params.gain.item(),
        pk1.params.Q.item(),
        pk2.params.freq.item(),
        pk2.params.gain.item(),
        pk2.params.Q.item(),
        ls.params.freq.item(),
        ls.params.gain.item(),
        hs.params.freq.item(),
        hs.params.gain.item(),
        lp.params.freq.item(),
        lp.params.Q.item(),
        hp.params.freq.item(),
        hp.params.Q.item(),
    ]
    update_fx_outputs = [
        pk1_freq,
        pk1_gain,
        pk1_q,
        pk2_freq,
        pk2_gain,
        pk2_q,
        ls_freq,
        ls_gain,
        hs_freq,
        hs_gain,
        lp_freq,
        lp_q,
        hp_freq,
        hp_q,
    ]
    update_plots = lambda: [
        plot_eq(),
        plot_comp(),
        plot_delay(),
        plot_reverb(),
        plot_t60(),
    ]
    update_plots_outputs = [
        peq_plot,
        comp_plot,
        delay_plot,
        reverb_plot,
        t60_plot,
    ]

    update_all = lambda i: update_pc(i) + update_fx() + update_plots()
    update_all_outputs = update_pc_outputs + update_fx_outputs + update_plots_outputs

    random_button.click(
        chain_functions(
            lambda i: (z.normal_(0, 1).clip_(SLIDER_MIN, SLIDER_MAX), i),
            lambda args: (z2fx(), args[1]),
            lambda args: args[1],
            update_all,
        ),
        inputs=extra_pc_dropdown,
        outputs=update_all_outputs,
    )
    reset_button.click(
        # lambda: (lambda _: [0 for _ in range(NUMBER_OF_PCS + 1)])(z.zero_()),
        lambda: chain_functions(
            lambda _: z.zero_(),
            lambda _: z2fx(),
            lambda _: update_all(NUMBER_OF_PCS),
        )(None),
        outputs=update_all_outputs,
    )

    def update_z(s, i):
        z[i] = s
        return

    for i, slider in enumerate(sliders):
        slider.input(
            chain_functions(
                partial(update_z, i=i),
                lambda _: z2fx(),
                lambda _: update_fx() + update_plots() + [model2json()],
            ),
            inputs=slider,
            outputs=update_fx_outputs + update_plots_outputs + [json_output],
        )
    extra_slider.input(
        lambda *xs: chain_functions(
            lambda args: update_z(args[0], args[1] - 1),
            lambda _: z2fx(),
            lambda _: update_fx() + update_plots() + [model2json()],
        )(xs),
        inputs=[extra_slider, extra_pc_dropdown],
        outputs=update_fx_outputs + update_plots_outputs + [json_output],
    )

    extra_pc_dropdown.input(
        lambda i: z[i - 1].item(),
        inputs=extra_pc_dropdown,
        outputs=extra_slider,
    )

demo.launch()