| """Checks quran_asr.features against torchaudio's Kaldi-compliance fbank (the reference recipe).""" |
| import os |
| import sys |
|
|
| import torch |
|
|
| sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) |
| from quran_asr.features import fbank |
|
|
|
|
| def main(): |
| import torchaudio.compliance.kaldi as kaldi |
|
|
| torch.manual_seed(0) |
| worst = 0.0 |
| for n in (16000, 16000 * 3 + 123, 48000 + 77, 5000): |
| wave = (torch.randn(n) * 0.1).clamp(-1, 1) |
| ours = fbank(wave) |
| ref = kaldi.fbank( |
| wave[None, :], num_mel_bins=80, frame_length=25.0, frame_shift=10.0, dither=0.0, |
| preemphasis_coefficient=0.97, remove_dc_offset=True, window_type="povey", |
| round_to_power_of_two=True, low_freq=20.0, high_freq=-400.0, snip_edges=False, |
| use_energy=False, use_log_fbank=True, use_power=True, sample_frequency=16000.0, |
| ) |
| assert ours.shape == ref.shape, (ours.shape, ref.shape) |
| diff = (ours - ref).abs().max().item() |
| worst = max(worst, diff) |
| print(f"{n} samples: {ours.shape[0]} frames, max |diff| = {diff:.2e}") |
| assert worst < 1e-3, worst |
| print("OK: features match the Kaldi recipe") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
