After experimenting with stft and istft, I have discovered that the input rank limitation imposed by torch.stft is inconvenient.
When containing multiple channels in single waveform, the input shape might be (batch_size, num_channels, num_samples) which is not supported by torch.stft. However, this format works well with other backends.
To address this, the PR has added reshape logic for stft function in the torch backend. The rank limitation of ops.stft has now been removed.
Furthermore, I have found that torch.istft works under specific condition which is common in audio processing.
The current implementation can be summarized in the following table:
After experimenting with
stftandistft, I have discovered that the input rank limitation imposed bytorch.stftis inconvenient.When containing multiple channels in single waveform, the input shape might be
(batch_size, num_channels, num_samples)which is not supported bytorch.stft. However, this format works well with other backends.To address this, the PR has added
reshapelogic forstftfunction in the torch backend. The rank limitation ofops.stfthas now been removed.Furthermore, I have found that
torch.istftworks under specific condition which is common in audio processing.The current implementation can be summarized in the following table:
scipy.signal.stftscipy.signal.istftjax.scipy.signal.stftjax.scipy.signal.istfttf.signal.stfttf.signal.inverse_stfttorch.stfttorch.istft& custom implementation