app.py

import gradio as gr
import io
import typing as T

import numpy as np
from PIL import Image
import pydub
from scipy.io import wavfile
import torch
import torchaudio

def convert(audio):
    # read uploaded file to wav
    rate, data = wavfile.read(audio)
    
    # resample from 48000 to 44100
    # from scipy.signal import resample
    # data = resample(data, int(data.shape[0] * 44100 / 48000))
    
    # convert to mono
    data = np.mean(data, axis=0)
    
    # convert to float32
    data = data.astype(np.float32)
    
    # take a random 7 second slice of the audio
    data = data[rate*7:rate*14]
    
    spectrogram = spectrogram_from_waveform(
        waveform=data,
        sample_rate=rate,
        # width=768,
        n_fft=8192,
        hop_length=512,
        win_length=8192,
    )

    spec = image_from_spectrogram(spectrogram)

    return spec

def spectrogram_from_waveform(
    waveform: np.ndarray,
    sample_rate: int,
    n_fft: int,
    hop_length: int,
    win_length: int,
    mel_scale: bool = True,
    n_mels: int = 512,
) -> np.ndarray:
    """
    Compute a spectrogram from a waveform.
    """

    spectrogram_func = torchaudio.transforms.Spectrogram(
        n_fft=n_fft,
        power=None,
        hop_length=hop_length,
        win_length=win_length,
    )

    waveform_tensor = torch.from_numpy(waveform.astype(np.float32)).reshape(1, -1)
    Sxx_complex = spectrogram_func(waveform_tensor).numpy()[0]

    Sxx_mag = np.abs(Sxx_complex)

    if mel_scale:
        mel_scaler = torchaudio.transforms.MelScale(
            n_mels=n_mels,
            sample_rate=sample_rate,
            f_min=0,
            f_max=10000,
            n_stft=n_fft // 2 + 1,
            norm=None,
            mel_scale="htk",
        )

        Sxx_mag = mel_scaler(torch.from_numpy(Sxx_mag)).numpy()

    return Sxx_mag

def image_from_spectrogram(
    spectrogram: np.ndarray, max_volume: float = 50, power_for_image: float = 0.25
) -> Image.Image:
    """
    Compute a spectrogram image from a spectrogram magnitude array.
    """
    # Apply the power curve
    data = np.power(spectrogram, power_for_image)

    # Rescale to 0-255
    data = data * 255 / max_volume

    # Invert
    data = 255 - data

    # Convert to a PIL image
    image = Image.fromarray(data.astype(np.uint8))

    # Flip Y
    image = image.transpose(Image.FLIP_TOP_BOTTOM)

    # Convert to RGB
    image = image.convert("RGB")

    return image

gr.Interface(fn=convert, inputs=[gr.Audio(source="upload", type="filepath")], outputs=[gr.Image()]).launch()