app.py

import math
import tempfile
from typing import Optional, Tuple, Union

import gradio
import gradio.inputs
import gradio.outputs
import markdown
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import torch
from loguru import logger
from torch import Tensor
from torchaudio.backend.common import AudioMetaData

from df import config
from df.enhance import enhance, init_df, load_audio, save_audio
from df.io import resample

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, df, _ = init_df("./DeepFilterNet2", config_allow_defaults=True)
model = model.to(device=device).eval()

fig_noisy: plt.Figure
fig_enh: plt.Figure
ax_noisy: plt.Axes
ax_enh: plt.Axes
fig_noisy, ax_noisy = plt.subplots(figsize=(15.2, 4))
fig_noisy.set_tight_layout(True)
fig_enh, ax_enh = plt.subplots(figsize=(15.2, 4))
fig_enh.set_tight_layout(True)

NOISES = {
    "None": None,
    "Kitchen": "samples/dkitchen.wav",
    "Living Room": "samples/dliving.wav",
    "River": "samples/nriver.wav",
    "Cafe": "samples/scafe.wav",
}


def mix_at_snr(clean, noise, snr, eps=1e-10):
    """Mix clean and noise signal at a given SNR.

    Args:
        clean: 1D Tensor with the clean signal to mix.
        noise: 1D Tensor of shape.
        snr: Signal to noise ratio.

    Returns:
        clean: 1D Tensor with gain changed according to the snr.
        noise: 1D Tensor with the combined noise channels.
        mix: 1D Tensor with added clean and noise signals.

    """
    clean = torch.as_tensor(clean).mean(0, keepdim=True)
    noise = torch.as_tensor(noise).mean(0, keepdim=True)
    if noise.shape[1] < clean.shape[1]:
        noise = noise.repeat((1, int(math.ceil(clean.shape[1] / noise.shape[1]))))
    max_start = int(noise.shape[1] - clean.shape[1])
    start = torch.randint(0, max_start, ()).item() if max_start > 0 else 0
    logger.debug(f"start: {start}, {clean.shape}")
    noise = noise[:, start : start + clean.shape[1]]
    E_speech = torch.mean(clean.pow(2)) + eps
    E_noise = torch.mean(noise.pow(2))
    K = torch.sqrt((E_noise / E_speech) * 10 ** (snr / 10) + eps)
    noise = noise / K
    mixture = clean + noise
    logger.debug("mixture: {mixture.shape}")
    assert torch.isfinite(mixture).all()
    max_m = mixture.abs().max()
    if max_m > 1:
        logger.warning(f"Clipping detected during mixing. Reducing gain by {1/max_m}")
        clean, noise, mixture = clean / max_m, noise / max_m, mixture / max_m
    return clean, noise, mixture


def load_audio_gradio(
    audio_or_file: Union[None, str, Tuple[int, np.ndarray]], sr: int
) -> Optional[Tuple[Tensor, AudioMetaData]]:
    if audio_or_file is None:
        return None
    if isinstance(audio_or_file, str):
        if audio_or_file.lower() == "none":
            return None
        # First try default format
        audio, meta = load_audio(audio_or_file, sr)
    else:
        meta = AudioMetaData(-1, -1, -1, -1, "")
        assert isinstance(audio_or_file, (tuple, list))
        meta.sample_rate, audio_np = audio_or_file
        # Gradio documentation says, the shape is [samples, 2], but apparently sometimes its not.
        audio_np = audio_np.reshape(audio_np.shape[0], -1).T
        if audio_np.dtype == np.int16:
            audio_np = (audio_np / (1 << 15)).astype(np.float32)
        elif audio_np.dtype == np.int32:
            audio_np = (audio_np / (1 << 31)).astype(np.float32)
        audio = resample(torch.from_numpy(audio_np), meta.sample_rate, sr)
    return audio, meta


def demo_fn(speech_upl: str, noise_type: str, snr: int):
    sr = config("sr", 48000, int, section="df")
    logger.info(f"Got parameters speech_upl: {speech_upl}, noise: {noise_type}, snr: {snr}")
    snr = int(snr)
    noise_fn = NOISES[noise_type]
    meta = AudioMetaData(-1, -1, -1, -1, "")
    max_s = 10  # limit to 10 seconds
    if speech_upl is not None:
        sample, meta = load_audio(speech_upl, sr)
        max_len = max_s * sr
        if sample.shape[-1] > max_len:
            start = torch.randint(0, sample.shape[-1] - max_len, ()).item()
            sample = sample[..., start : start + max_len]
    else:
        sample, meta = load_audio("samples/p232_013_clean.wav", sr)
        sample = sample[..., : max_s * sr]
    if sample.dim() > 1 and sample.shape[0] > 1:
        assert (
            sample.shape[1] > sample.shape[0]
        ), f"Expecting channels first, but got {sample.shape}"
        sample = sample.mean(dim=0, keepdim=True)
    logger.info(f"Loaded sample with shape {sample.shape}")
    if noise_fn is not None:
        noise, _ = load_audio(noise_fn, sr)  # type: ignore
        logger.info(f"Loaded noise with shape {noise.shape}")
        _, _, sample = mix_at_snr(sample, noise, snr)
    logger.info("Start denoising audio")
    enhanced = enhance(model, df, sample)
    logger.info("Denoising finished")
    lim = torch.linspace(0.0, 1.0, int(sr * 0.15)).unsqueeze(0)
    lim = torch.cat((lim, torch.ones(1, enhanced.shape[1] - lim.shape[1])), dim=1)
    enhanced = enhanced * lim
    if meta.sample_rate != sr:
        enhanced = resample(enhanced, sr, meta.sample_rate)
        sample = resample(sample, sr, meta.sample_rate)
        sr = meta.sample_rate
    noisy_fn = tempfile.NamedTemporaryFile(suffix="noisy.wav", delete=False).name
    save_audio(noisy_fn, sample, sr)
    enhanced_fn = tempfile.NamedTemporaryFile(suffix="enhanced.wav", delete=False).name
    save_audio(enhanced_fn, enhanced, sr)
    logger.info(f"saved audios: {noisy_fn}, {enhanced_fn}")
    ax_noisy.clear()
    ax_enh.clear()
    return (
        gradio.make_waveform(noisy_fn, bar_count=200),
        spec_im(sample, sr=sr, figure=fig_noisy, ax=ax_noisy),
        gradio.make_waveform(enhanced_fn, bar_count=200),
        spec_im(enhanced, sr=sr, figure=fig_enh, ax=ax_enh),
    )


def specshow(
    spec,
    ax=None,
    title=None,
    xlabel=None,
    ylabel=None,
    sr=48000,
    n_fft=None,
    hop=None,
    t=None,
    f=None,
    vmin=-100,
    vmax=0,
    xlim=None,
    ylim=None,
    cmap="inferno",
):
    """Plots a spectrogram of shape [F, T]"""
    spec_np = spec.cpu().numpy() if isinstance(spec, torch.Tensor) else spec
    if ax is not None:
        set_title = ax.set_title
        set_xlabel = ax.set_xlabel
        set_ylabel = ax.set_ylabel
        set_xlim = ax.set_xlim
        set_ylim = ax.set_ylim
    else:
        ax = plt
        set_title = plt.title
        set_xlabel = plt.xlabel
        set_ylabel = plt.ylabel
        set_xlim = plt.xlim
        set_ylim = plt.ylim
    if n_fft is None:
        if spec.shape[0] % 2 == 0:
            n_fft = spec.shape[0] * 2
        else:
            n_fft = (spec.shape[0] - 1) * 2
    hop = hop or n_fft // 4
    if t is None:
        t = np.arange(0, spec_np.shape[-1]) * hop / sr
    if f is None:
        f = np.arange(0, spec_np.shape[0]) * sr // 2 / (n_fft // 2) / 1000
    im = ax.pcolormesh(
        t, f, spec_np, rasterized=True, shading="auto", vmin=vmin, vmax=vmax, cmap=cmap
    )
    if title is not None:
        set_title(title)
    if xlabel is not None:
        set_xlabel(xlabel)
    if ylabel is not None:
        set_ylabel(ylabel)
    if xlim is not None:
        set_xlim(xlim)
    if ylim is not None:
        set_ylim(ylim)
    return im


def spec_im(
    audio: torch.Tensor,
    figsize=(15, 5),
    colorbar=False,
    colorbar_format=None,
    figure=None,
    labels=True,
    **kwargs,
) -> Image:
    audio = torch.as_tensor(audio)
    if labels:
        kwargs.setdefault("xlabel", "Time [s]")
        kwargs.setdefault("ylabel", "Frequency [Hz]")
    n_fft = kwargs.setdefault("n_fft", 1024)
    hop = kwargs.setdefault("hop", 512)
    w = torch.hann_window(n_fft, device=audio.device)
    spec = torch.stft(audio, n_fft, hop, window=w, return_complex=False)
    spec = spec.div_(w.pow(2).sum())
    spec = torch.view_as_complex(spec).abs().clamp_min(1e-12).log10().mul(10)
    kwargs.setdefault("vmax", max(0.0, spec.max().item()))

    if figure is None:
        figure = plt.figure(figsize=figsize)
        figure.set_tight_layout(True)
    if spec.dim() > 2:
        spec = spec.squeeze(0)
    im = specshow(spec, **kwargs)
    if colorbar:
        ckwargs = {}
        if "ax" in kwargs:
            if colorbar_format is None:
                if kwargs.get("vmin", None) is not None or kwargs.get("vmax", None) is not None:
                    colorbar_format = "%+2.0f dB"
            ckwargs = {"ax": kwargs["ax"]}
        plt.colorbar(im, format=colorbar_format, **ckwargs)
    figure.canvas.draw()
    return Image.frombytes("RGB", figure.canvas.get_width_height(), figure.canvas.tostring_rgb())


inputs = [
    gradio.Audio(
        label="Upload audio sample",
        source="upload",
        type="filepath",
    ),
    gradio.Dropdown(
        label="Add background noise",
        choices=list(NOISES.keys()),
        value="None",
    ),
    gradio.Dropdown(
        label="Noise Level (SNR)",
        choices=["-5", "0", "10", "20"],
        value="10",
    ),
]
outputs = [
    gradio.Video(type="filepath", label="Noisy audio"),
    gradio.Image(label="Noisy spectrogram"),
    gradio.Video(type="filepath", label="Noisy audio"),
    gradio.Image(label="Enhanced spectrogram"),
]
description = "This demo denoises audio files using DeepFilterNet. Try it with your own voice!"
iface = gradio.Interface(
    fn=demo_fn,
    title="DeepFilterNet2 Demo",
    inputs=inputs,
    outputs=outputs,
    description=description,
    allow_flagging="never",
    article=markdown.markdown(open("usage.md").read()),
)
iface.launch(debug=True)