app.py

import glob
import math
import os
import tempfile
import time
from typing import List, Optional, Tuple, Union

import subprocess
# import os

import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import torch
from loguru import logger
from PIL import Image
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

# Optimize PyTorch CPU execution for faster inference
if torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")
    # Limit intra-op thread count to avoid scheduling overhead on multi-core environments
    torch.set_num_threads(4)

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 ensure_wav(filepath: str) -> str:
    """Convert MP3 (or other non-WAV formats) to WAV using ffmpeg if needed."""
    if not filepath.lower().endswith(".wav"):
        wav_path = filepath.rsplit(".", 1)[0] + ".wav"
        if wav_path == filepath:
            wav_path = filepath + ".wav"
        # Run ffmpeg silently to convert the format
        subprocess.run(
            ["ffmpeg", "-y", "-i", filepath, wav_path],
            check=True,
            stdout=subprocess.DEVNULL,
            stderr=subprocess.DEVNULL
        )
        return wav_path
    return filepath


def wav_to_mp3(wav_path: str) -> str:
    """Convert temporary WAV output to MP3 using ffmpeg for smaller file size."""
    mp3_path = wav_path.rsplit(".", 1)[0] + ".mp3"
    subprocess.run(
        ["ffmpeg", "-y", "-i", wav_path, "-vn", "-ar", "44100", "-ac", "1", "-b:a", "128k", mp3_path],
        check=True,
        stdout=subprocess.DEVNULL,
        stderr=subprocess.DEVNULL
    )
    # Remove the temporary wav file
    try:
        os.remove(wav_path)
    except Exception as e:
        logger.warning(f"Failed to remove temp WAV file {wav_path}: {e}")
    return mp3_path


import torchaudio.functional as F_audio

def post_process_audio(waveform: Tensor, sr: int) -> Tensor:
    """Apply professional DSP filters to make voice audio crisp and clear."""
    # 1. High-pass filter at 80 Hz to cut muddy sub-bass rumble and AC hums
    waveform = F_audio.highpass_biquad(waveform, sample_rate=sr, cutoff_freq=80.0)

    # 2. Treble peaking equalizer at 6000 Hz (+3.0 dB) to add air, presence, and vocal crispness
    waveform = F_audio.equalizer_biquad(waveform, sample_rate=sr, center_freq=6000.0, gain=3.0, Q=0.707)

    # 3. Normalize peak to -1.0 dBFS (0.9 max amplitude) to maximize volume without clipping
    max_val = waveform.abs().max()
    if max_val > 0:
        waveform = waveform * (0.9 / max_val)

    return waveform


def demo_fn(
    speech_upl: str,
    noise_type: str,
    snr: int,
    atten_lim_db: float,
    wet_dry_mix: float,
    post_process: bool,
):

    sr = config("sr", 48000, int, section="df")
    logger.info(f"Got parameters speech_upl: {speech_upl}, noise: {noise_type}, snr: {snr}, atten_lim_db: {atten_lim_db}, wet_dry_mix: {wet_dry_mix}, post_process: {post_process}")
    snr = int(snr)
    noise_fn = NOISES[noise_type]
    meta = AudioMetaData(-1, -1, -1, -1, "")

    max_s = 3600  # allow up to 1 hour (3600 seconds)

    if speech_upl is not None:
        # ✅ Ensure compatible WAV input
        speech_upl = ensure_wav(speech_upl)

        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")
    # Call enhance with attenuation limit (atten_lim_db) to prevent over-silencing & artifacts
    with torch.inference_mode():
        enhanced = enhance(model, df, sample, atten_lim_db=atten_lim_db)
    logger.info("Denoising finished")

    # Dry/wet blending: mix original 'sample' back into 'enhanced' to restore voice texture
    # Ensure both tensors have the exact same shape (DeepFilterNet can output a slightly different length due to STFT framing)
    min_len = min(enhanced.shape[-1], sample.shape[-1])
    enhanced = enhanced[..., :min_len]
    sample = sample[..., :min_len]

    alpha = wet_dry_mix / 100.0
    enhanced = alpha * enhanced + (1 - alpha) * sample

    # Apply professional post-processing (Low-cut filter + presence boost EQ)
    if post_process:
        enhanced = post_process_audio(enhanced, sr)

    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_wav = tempfile.NamedTemporaryFile(suffix="noisy.wav", delete=False).name
    save_audio(noisy_wav, sample, sr)
    enhanced_wav = tempfile.NamedTemporaryFile(suffix="enhanced.wav", delete=False).name
    save_audio(enhanced_wav, enhanced, sr)

    logger.info("Converting outputs to MP3...")
    noisy_mp3 = wav_to_mp3(noisy_wav)
    enhanced_mp3 = wav_to_mp3(enhanced_wav)

    logger.info(f"saved audios: {noisy_mp3}, {enhanced_mp3}")

    ax_noisy.clear()
    ax_enh.clear()
    noisy_im = spec_im(sample, sr=sr, figure=fig_noisy, ax=ax_noisy)
    enh_im = spec_im(enhanced, sr=sr, figure=fig_enh, ax=ax_enh)

    filter = [speech_upl, noisy_mp3, enhanced_mp3]
    cleanup_tmp(filter)

    return noisy_mp3, noisy_im, enhanced_mp3, enh_im

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())


def cleanup_tmp(filter: List[str] = [], hours_keep=2):
    filter.append("p232")
    logger.info(f"Filter: {filter}")
    # Cleanup some old wav files
    if os.path.exists("/tmp"):
        for f in glob.glob("/tmp/*"):
            print(f"Got file {f}")
            is_old = (time.time() - os.path.getmtime(f)) / 3600 > hours_keep
            filtered = any(filt in f for filt in filter if filt is not None)
            if is_old and not filtered:
                try:
                    os.remove(f)
                    logger.info(f"Removed file {f}")
                except Exception as e:
                    logger.warning(f"failed to remove file {f}: {e}")
with gr.Blocks() as demo:
    with gr.Row():
        gr.Markdown(
            """
            ## DeepFilterNet2 Demo\

            This demo denoises audio files using DeepFilterNet. Try it with your own voice!
            """
        )
    with gr.Row():
        with gr.Column():
            audio_file = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Input Audio")
            inputs = [
                audio_file,
                gr.Dropdown(
                    label="Add background noise",
                    choices=list(NOISES.keys()),
                    value="None",
                ),
                gr.Dropdown(
                    label="Noise Level (SNR)",
                    choices=["-5", "0", "10", "20"],
                    value="10",
                ),
                gr.Slider(
                    minimum=6,
                    maximum=35,
                    step=1,
                    value=15,
                    label="Max Attenuation (dB) - lower is more natural/clear, higher reduces more noise",
                ),
                gr.Slider(
                    minimum=50,
                    maximum=100,
                    step=5,
                    value=90,
                    label="Voice Naturalness Mix (%) - 100% is fully denoised, 90% blends back some original voice",
                ),
                gr.Checkbox(
                    label="Post-Process (80Hz Low-Cut & Presence Boost)",
                    value=True,
                ),
            ]
            btn = gr.Button("Generate")
        with gr.Column():
            outputs = [
                # gr.Video(type="filepath", label="Noisy audio"),
                gr.Audio(type="filepath", label="Noisy audio"),
                gr.Image(label="Noisy spectrogram"),
                # gr.Video(type="filepath", label="Enhanced audio"),
                gr.Audio(type="filepath", label="Enhanced audio"),
                gr.Image(label="Enhanced spectrogram"),
            ]
    btn.click(fn=demo_fn, inputs=inputs, outputs=outputs, api_name='denoise')
    gr.Examples(
        [
            ["./samples/p232_013_clean.wav", "Kitchen", "10", 15, 90, True],
            ["./samples/p232_013_clean.wav", "Cafe", "10", 15, 90, True],
            ["./samples/p232_019_clean.wav", "Cafe", "10", 15, 90, True],
            ["./samples/p232_019_clean.wav", "River", "10", 15, 90, True],
        ],
        fn=demo_fn,
        inputs=inputs,
        outputs=outputs,
        cache_examples=True,
    )
    gr.Markdown(open("usage.md").read())

cleanup_tmp()
# demo.launch(enable_queue=True)
# demo.launch()
demo.queue().launch()