Visualize mel difference

diff_melspec_16khz_vs_22khz.py

wav_path = str(REPO_PATH / "tests/accent_test_sample/split/accent17/accent17_a17_14s.wav")
wav_16, sr16 = librosa.load(wav_path, sr=16000)
wav_22, sr22 = librosa.load(wav_path, sr=22000)
print(wav_22.shape, wav_16.shape)

# test1: sr=16khz, win=640, hop=160
from new_data.preprocessing.batch_processors.spectrogram_creator import (
    SpectrogramCreator
)
from new_data.train_data_containers import Batch
mel_fn = SpectrogramCreator(
    sample_rate=16000,
    filter_length=640,
    win_length=640,
    hop_length=160,
    device="cpu",
)
res16 = mel_fn(Batch(
    audio=wav_16[None],
    audio_len=torch.tensor([len(wav_16)]),
))

# test2: sr=22khz, win=880, hop=220
from new_data.preprocessing.batch_processors.spectrogram_creator import (
    SpectrogramCreator
)
from new_data.train_data_containers import Batch
mel_fn = SpectrogramCreator(
    sample_rate=22000,
    filter_length=880,
    win_length=880,
    hop_length=220,
    device="cpu",
)
res22 = mel_fn(Batch(
    audio=wav_22[None],
    audio_len=torch.tensor([len(wav_22)]),
))

# visualize
print(res16['spectrogram'].shape)
mel16, mel22 = (b['spectrogram'][:, :, :78] for b in (res16, res22))

print(mel16.shape, mel16.min(), mel16.max())
print(mel22.shape, mel22.min(), mel22.max())

# I perform linear transform first
mel22 = (mel22 - mel22.min()) / (mel22.max() - mel22.min()) # to [0, 1]
mel22 = mel22 * (mel16.max() - mel16.min()) + mel16.min()

from utils.plot import plot_output, calc_shift, plot_shift_errors, plot_error
plot_output(mel16, mel22, f=None)
plt.show()

shifts = list(range(60))
shift, diffs, errors = calc_shift(mel16, mel22, shifts)
print("calced shift = ", shift)
print("with MAE=", errors[shift].item())
plot_shift_errors(shifts, errors, f=None)
plt.show()

plot_error(mel16, mel22, shift, f=None)
plt.show()