The spectrogram computed by "torchaudio.compliance.kaldi.spectrogram" and "compute-spectrogram-feats" are different

[yjw123456]

Dear Sir or Madam,

I compared the spectrogram generated by "torchaudio.compliance.kaldi.spectrogram" and "compute-spectrogram-feats" using default settings, however the outputs are different.

# The code for torchaudio is: 
wav, sample_rate = torchaudio.load('wav/1.wav')

spectrum=torchaudio.compliance.kaldi.spectrogram(wav, blackman_coeff=0.42, channel=-1, dither=1.0, energy_floor=0.0, frame_length=25.0, frame_shift=10.0, min_duration=0.0, preemphasis_coefficient=0.97, raw_energy=True, remove_dc_offset=True, round_to_power_of_two=True, sample_frequency=16000.0, snip_edges=True, subtract_mean=False, window_type='povey')

# Kaldi is:
compute-spectrogram-feats  --blackman_coeff=0.42 --channel=-1 --dither=1.0 --energy_floor=0.0 --frame_length=25.0 --frame_shift=10.0 --min_duration=0.0 --preemphasis_coefficient=0.97 --raw_energy=True --remove_dc_offset=True --round_to_power_of_two=True --sample_frequency=16000.0 --snip_edges=True --subtract_mean=False --window_type='povey'   scp:wav.scp ark,t,scp:1_stft.ark,1_stft.scp

Can you give me some clue of it?

Thanks!

[wuqiangch]

@yjw123456 I have the same problem.

[wuqiangch]

@yjw123456 I have the same problem.

[vincentqb]

The parameter given for sample_frequency to torchaudio.compliance.kaldi.spectrogram must be the sample rate for the audio file, sample_rate. Can you confirm that sample_rate is 16000 ?

Can you provide the whole code for kaldi so I can reproduce?

Here's an audio file to try with.

[yjw123456]

Hi, vincentqb,

Sorry for the late reply. @vincentqb @wuqiangch

I found the answer of this problem.
The waveform loaded by torchaudio is transformed into float type. While kaldi directly uses integers to calculate the spectrogram.

So if I do:
wav, sample_rate = torchaudio.load('tmp.wav')
wav = wav*2**15
Then the outputs will be exactly the same.

The whole code in kaldi is:
compute-spectrogram-feats --blackman_coeff=0.42 --channel=-1 --dither=1.0 --energy_floor=0.0 --frame_length=25.0 --frame_shift=10.0 --min_duration=0.0 --preemphasis_coefficient=0.97 --raw_energy=True --remove_dc_offset=True --round_to_power_of_two=True --sample_frequency=16000.0 --snip_edges=True --subtract_mean=False --window_type='povey' scp:wav.scp ark,t,scp:1_stft.ark,1_stft.scp

In the wav.scp is:
utt-id tmp.wav

[vincentqb]

Glad this answers your question. Yes, torchaudio normalizes waveforms to [-1, 1]. I will close this issue, but please feel free to reopen or create a new one.