app.py

import gradio
import gradio.inputs
import gradio.outputs
import torch
from df.enhance import enhance, init_df

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


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)
    noise = torch.as_tensor(noise)
    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
    assert torch.isfinite(mixture)
    return clean, noise, mixture


def mix_and_denoise(speech, noise, snr):
    model, df, _ = init_df()
    speech, noise, noisy = mix_at_snr(speech, noise, snr)
    enhanced = enhance(model.to(device=device).eval(), df, noisy)
    return speech, noisy, enhanced


inputs = [
    gradio.inputs.Audio(
        source="microphone", type="filepath", optional=True, label="Speech"
    ),
    gradio.inputs.Audio(
        source="microphone", type="filepath", optional=True, label="Noise"
    ),
    gradio.inputs.Slider(minimum=-10, maximum=40, step=5, default=10),
]
examples = [
    [],
    ["samples/noise_freesound_2530.wav", "samples/noise_freesound_573577.wav"],
]
outputs = [
    gradio.outputs.Audio(label="Clean"),
    gradio.outputs.Audio(label="Noisy"),
    gradio.outputs.Audio(label="Enhanced"),
]
iface = gradio.Interface(
    fn=mix_and_denoise, inputs=inputs, outputs=outputs, examples=examples
)
iface.launch()