app.py

from __future__ import annotations

import gradio as gr
try:
    import spaces
except ImportError:  # 'spaces' is only provided by Hugging Face Spaces
    import types as _types

    def _gpu(*args, **kwargs):
        if len(args) == 1 and callable(args[0]) and not kwargs:
            return args[0]

        def _decorator(func):
            return func

        return _decorator

    spaces = _types.SimpleNamespace(GPU=_gpu)

from pyharp import *


import tempfile
from voicefixer import VoiceFixer

# Initialize VoiceFixer. It handles downloading checkpoints and setting device.
# The VoiceFixer class automatically detects and uses CUDA if available.
voicefixer_model = VoiceFixer()


model_card = ModelCard(
    name="VoiceFixer",
    description=(
        "VoiceFixer aims to restore human speech regardless how serious its "
        "degraded. It can handle noise, reverberation, low resolution "
        "(2kHz~44.1kHz) and clipping (0.1-1.0 threshold) effect within one "
        "model. The model is fixed-rate and mono, so the result is resampled "
        "back to the input's rate and copied across its channels: the output "
        "has the format it came in with, but no stereo image. Repository: "
        "https://github.com/haoheliu/voicefixer Paper: "
        "https://arxiv.org/abs/2109.13731"
    ),
    author="Haohe Liu, Qiuqiang Kong, Qiao Tian, Yan Zhao, DeLiang Wang, Chuanzeng Huang, Yuxuan Wang",
    tags=[Subcategory.RESTORATION, SampleRate(44100), Channels(1), 'speech'],
)


@spaces.GPU
def match_input_format(output_path: str, input_path: str) -> str:
    """
    Rewrite an output file at the sample rate and channel count it came in with.

    The models here are fixed-rate and mono, and hand the writing to vendored
    upstream code, so the format is corrected afterwards rather than threaded
    through inference. Duplicated channels are copies, not a recovered stereo
    image; the model card says so.

    Args:
        output_path (str): File the model wrote. Overwritten in place.
        input_path (str): File the request supplied.

    Returns:
        output_path (str): The same path, for use in a return statement.
    """

    import librosa
    import numpy as np
    import soundfile

    want = soundfile.info(input_path)
    have = soundfile.info(output_path)

    if have.samplerate == want.samplerate and have.channels == want.channels:
        return output_path

    data, rate = soundfile.read(output_path, always_2d=True)

    if rate != want.samplerate:
        data = librosa.resample(
            data.T, orig_sr=rate, target_sr=want.samplerate
        ).T

    if data.shape[1] == 1 and want.channels > 1:
        data = np.tile(data, (1, want.channels))
    elif data.shape[1] > 1 and want.channels == 1:
        data = data.mean(axis=1, keepdims=True)

    soundfile.write(output_path, data, want.samplerate)

    return output_path


def process_fn(input_audio, mode):
    output_file = tempfile.NamedTemporaryFile(suffix=".wav", delete=False).name
    voicefixer_model.restore(input_audio, output_file, mode=int(mode))

    return match_input_format(output_file, input_audio)


if __name__ == "__main__":
    with gr.Blocks() as demo:
        input_components = [
            gr.Audio(type="filepath", label="Input Audio").harp_required(True).set_info("Upload an audio file to be processed by VoiceFixer."),
            gr.Dropdown(choices=["0", "1", "2"], value="0", label="Processing Mode", info="Select the VoiceFixer processing mode:\n0: Original Model (suggested by default)\n1: Add preprocessing module (remove higher frequency)\n2: Train mode (might work sometimes on seriously degraded real speech)"),
        ]
        output_components = [
            gr.Audio(type="filepath", label="Fixed Audio").set_info("The enhanced audio output from VoiceFixer."),
        ]
        build_endpoint(
            model_card=model_card,
            input_components=input_components,
            output_components=output_components,
            process_fn=process_fn,
        )

    demo.queue().launch(show_error=True, pwa=True)