test mel sp

test_mel_sp.py

import numpy as np
import librosa

"""
https://github.com/r9y9/wavenet_vocoder/blob/a835a8eebfec5e049d93f629c4ee2f1daab56329/train.py#L460
https://github.com/Rayhane-mamah/Tacotron-2/blob/ab5cb08a931fc842d3892ebeb27c8b8734ddd4b8/wavenet_vocoder/feeder.py#L368
https://github.com/kokeshing/WaveNet-Estimator/blob/cc2ee6bb699cf977356b23e7513ba549cdfc874f/dataset.py#L38
"""


def load_wav(path, sampling_rate):
    wav = librosa.core.load(path, sr=sampling_rate)[0]

    return wav


def get_mel_filter(sampling_rate, n_fft, num_mels, fmin=55, fmax=7600):
    mel_filter = librosa.filters.mel(sampling_rate, n_fft,
                                     n_mels=num_mels,
                                     fmin=fmin, fmax=fmax)

    return mel_filter


def melspectrogram(wav, mel_filter, n_fft=2048, hop_size=275,
                   win_size=1100, min_level_db=-100, ref_level_db=20):
    d = librosa.stft(y=wav, n_fft=n_fft, hop_length=hop_size,
                     win_length=win_size, pad_mode='constant')
    mel_sp = _linear_to_mel(np.abs(d), mel_filter)
    mel_sp = _amp_to_db(mel_sp, min_level_db) - ref_level_db

    return mel_sp


def _linear_to_mel(spectogram, mel_filter):
    return np.dot(mel_filter, spectogram)


def _amp_to_db(x, min_level_db):
    min_level = np.exp(min_level_db / 20 * np.log(10))

    return 20 * np.log10(np.maximum(min_level, x))


def main():
    sampling_rate = 22050

    mel_filter = get_mel_filter(sampling_rate, n_fft=2048, num_mels=80)

    all_wav = load_wav("test.wav", sampling_rate)
    all_mel_sp = melspectrogram(all_wav, mel_filter)
    mel_sp_len = all_mel_sp.shape[1]

    pad = (all_wav.shape[0] // 275 + 1) * 275 - all_wav.shape[0]
    all_wav = np.pad(all_wav, (0, pad), mode='constant', constant_values=0.0)
    all_wav = all_wav[:mel_sp_len * 275]

    max_time_frames = 40
    max_steps = 275 * 40
    for mel_offset in range(mel_sp_len - max_time_frames):
        wav_offset = mel_offset * 275
        wav = all_wav[wav_offset:wav_offset + max_steps]
        print(mel_offset)
        assert np.allclose(melspectrogram(wav, mel_filter)[:, :40], all_mel_sp[:, mel_offset:mel_offset + max_time_frames]), "not same mel sp"


if __name__ == '__main__':
    main()