A few questions... about lhotse

[pzelasko]

https://github.com/lhotse-speech/lhotse/blob/master/lhotse/kaldi.py#L68
why duration - start rather than just duration?

My bad - the variable should be called end, the code should work fine otherwise.

https://github.com/lhotse-speech/lhotse/blob/master/lhotse/audio.py#L178
why not [n_sources, n_channels, n_samples]?

I think that might be a naming collision - "source" in this context is supposed to mean "some storage containing audio samples for one or more channels" (e.g. a file), rather than an acoustic sound source in some physical space... So if some corpus has recordings from an array of 4 mics stored in separate wav files, there would be 4 AudioSource objects inside a Recording to represent that. Does that make it clearer?

I'll make a PR soon with updates to fix/clarify based on your feedback - thanks!

[freewym]

https://github.com/lhotse-speech/lhotse/blob/master/lhotse/kaldi.py#L68
why duration - start rather than just duration?

My bad - the variable should be called end, the code should work fine otherwise.

https://github.com/lhotse-speech/lhotse/blob/master/lhotse/audio.py#L178
why not [n_sources, n_channels, n_samples]?

I think that might be a naming collision - "source" in this context is supposed to mean "some storage containing audio samples for one or more channels" (e.g. a file), rather than an acoustic sound source in some physical space... So if some corpus has recordings from an array of 4 mics stored in separate wav files, there would be 4 AudioSource objects inside a Recording to represent that. Does that make it clearer?

I'll make a PR soon with updates to fix/clarify based on your feedback - thanks!

Thanks! what is the size of each element in the list samples_per_source? I assume it is already 2D?

[pzelasko]

Thanks! what is the size of each element in the list samples_per_source? I assume it is already 2D?

Yes - for mono files it'd be (1, num_samples), for stereo files (2, num_samples), and the stack operation concats across the channel dim.

[freewym]

I think np.vstack() will add one more dimension on top on the existing tensor, so the resulting tensor will be in 3D?

[pzelasko]

I don't think so:

import numpy
x = numpy.ones((1, 1000))
y = numpy.zeros((2, 1000))
numpy.vstack([x, y]).shape
Out[5]: (3, 1000)

[freewym]

Oh OK, maybe only for 1D vectors it adds one more dimension.

[csukuangfj]

I think np.vstack() will add one more dimension on top on the existing tensor

np.stack will add an extra dimension, not np.vstack.

[danpovey]

np.vstack seems to have a very confusing interface. If np.stack is usable, I'd prefer it.

[danpovey]

.. or np.concatenate if not, maybe?