8kHz audio

[patrickltobing]

Hi @JeremyCCHsu,

Could you successfully perform analysis-synthesis with WORLD on 8kHz audio?

[patrickltobing]

I have found the problem and quick solution.

The aperiodicity spectrum is failed if extracted from 8kHz.

The analysis-synthesis of 8kHz will still be working fine if the aperiodicity is extracted with 16kHz waveform,
the rests, spectral envelope and f0 are extracted from 8kHz.

[JeremyCCHsu]

Issue 1:

Hey @patrickltobing , thanks for the issue and the workaround. I confirmed it on my side, too. I'll consult the author of WORLD and get back to you.

Workaround:

wav_8k = librosa.resample(wav16k, fs_16k, fs_8k)

_, _, ap = pw.wav2world(wav16k, fs_16k)
f0, sp, _ = pw.wav2world(wav_8k, fs_8k)
ap = ap[:, :sp.shape[1]].copy(order="C")
y = pw.synthesize(f0, sp, ap, fs_8k)
sf.write(f"case2-using-16k-ap.wav", y, fs_8k)

Issue 2:

In addition to the issue you mentioned, I also found that DIO could sometimes yield suboptimal pitch contour with 16 kHz speech. In case you also have this issue use, pw.harvest instead of pw.dio should solve it. Again, I'm consulting the author for further information.

Workaround

_f0, t = pw.harvest(x, target_fs)    # raw pitch extractor
f0 = pw.stonemask(x, _f0, t, target_fs)  # pitch refinement
sp = pw.cheaptrick(x, f0, t, target_fs)  # extract smoothed spectrogram
ap = pw.d4c(x, f0, t, target_fs)         # extract aperiodicity