Add .narrowband() effect (mulaw, lpc10 codecs)

[rouseabout]

This patch adds a audio codec transformation.

I have found that when applying K2 ASR to speech compressed with mulaw, it is advantageous to augment the training data with these codecs. The transformation resamples the input audio to 8kHz, encodes then decodes using specified codec, then restores the original sample rate (e.g. 16 kHz).

Open issues:

• The transformation is called phone(). But maybe a better name is needed?
• Since it significantly alters the audio, depending on codec, I am wondering how best to test the transformation?

Example use:

cs2 = CutSet.from_manifests(...).phone(codec="mulaw")
cs3 = CutSet.from_manifests(...).phone(codec="lpc10")

libspandsp is required to use the lpc10 codec. Use apt-get install libspandsp-dev on Debian/Ubuntu.

[pzelasko]

Looks pretty good! I left a few comments. Could you also add unit tests for this transform?

[pzelasko]

I suggest calling it Narrowband and renaming the methods to narrowband. Also upsampling back to original SR should be optional (restore_orig_sr=True).

[pzelasko]

We shouldn't need this option at all. You can call get_or_create_resampler directly in __call__ using the input example's actual sampling rate. This way this transform can work with datasets of mixed sampling rates.

[rouseabout]

I have addressed everything except for restore_orig_sr=True. I am not sure how to achieve that!

[pzelasko]

I have addressed everything except for restore_orig_sr=True. I am not sure how to achieve that!

You are very close! Add a parameter restore_orig_sr=True in def narrowband(self, ...) for cut and recording, and pass the provided argument to Narrowband constructor. Then you can extend the condition for the second resampling to if self.restore_orig_sr and sampling_rate != 8000).

[rouseabout]

Done, but something extra is needed, because when I apply the transformation with use_orig_sr=False the following exception occurs:

AudioLoadingError: The number of declared samples in the recording diverged from the one obtained when loading audio (offset=0, duration=19.22419501133787). This could be internal Lhotse's error or a faulty transform implementation. Please report this issue in Lhotse and show the following: diff=693887, audio.shape=(1, 153900), recording=Recording(id='0_nb_lpc10', sources=[AudioSource(type='file', channels=[0], source='/home/user/workspace/rtvalid/0.wav')], sampling_rate=44100, num_samples=847787, duration=19.22419501133787, channel_ids=[0], transforms=[{'name': 'Narrowband', 'kwargs': {'codec': 'lpc10', 'restore_orig_sr': False}}])

[pzelasko]

If you don't restore orig sr, you'll have to update both sampling_rate and num_samples property on the Recording object.

[pzelasko]

Thanks for the contribution, merging!