mel spectrogram image to audio file

img2audio.py

import librosa
import numpy as np
from PIL import Image
import soundfile as sf


def image_to_audio(image, sr=22050, n_fft=2048, hop_length=512, top_db=80):
    """Converts spectrogram to audio.

    Args:
        image (PIL Image): x_res x y_res grayscale image

    Returns:
        audio (array): raw audio
    """
    bytedata = np.frombuffer(image.tobytes(), dtype="uint8").reshape(
        (image.width, image.height)
    )
    log_S = bytedata.astype("float") * top_db / 255 - top_db
    S = librosa.db_to_power(log_S)
    audio = librosa.feature.inverse.mel_to_audio(
        S, sr=sr, n_fft=n_fft, hop_length=hop_length
    )
    return audio


image = Image.open("./unnamed.png").convert("L")
if image.size != (256, 256):
    image = image.resize((256, 256))
    print("Image resized")
sound = image_to_audio(image)
sf.write("example.wav", sound, 22050)
print("Audio file saved")