Why does `transforms.TimeStretch` return `complex64`?

kuraga · November 3, 2023, 9:17pm

Good day!

torchaudio 2.1.1

TimeStretch — Torchaudio 2.1.0 documentation :

Stretched spectrogram. The resulting tensor is of the same dtype as the input spectrogram, but the number of frames is changed to ceil(num_frame / rate).

s = torchaudio.transforms.Spectrogram()(x)
s.dtype  # => torch.float32

t = torchaudio.transforms.TimeStretch(fixed_rate=0.9)(s)
t.dtype  # =>  torch.complex64

Why?

kuraga · November 9, 2023, 1:03am

github.com/pytorch/audio

Why does `transforms.TimeStretch` return of type `complex64`?

opened 12:02PM - 05 Nov 23 UTC

kuraga

### 🐛 Describe the bug Good day! https://pytorch.org/audio/2.1.0/generated/t…orchaudio.transforms.TimeStretch.html#torchaudio.transforms.TimeStretch.forward: > Stretched spectrogram. The resulting tensor is of the same dtype as the input spectrogram, but the number of frames is changed to `ceil(num_frame / rate)`. But: ``` s = torchaudio.transforms.Spectrogram()(x) s.dtype # => torch.float32 t = torchaudio.transforms.TimeStretch(fixed_rate=0.9)(s) t.dtype # => torch.complex64 ``` Should I collect a bug report or don't I understand time stretching? (previously posted [at the forum](https://discuss.pytorch.org/t/why-does-transforms-timestretch-return-complex64/191208)) ### Versions torchaudio 2.1.1 from Google Colab

Greg_Shay · June 19, 2024, 1:17pm

If you haven’t figured this out already…
A complex spectrum contains the phase as well as amplitude of the audio spectral energies.

If your application results are ok with ignoring phase, convert complex to real by using a magnitude function.

If you don’t want to lose the phase information, you will have to keep (and use) the complex spectrum results. Each frequency bin in the spectrum will have a phase according to the ratio of the real vs imaginary value of that bin.