Improve audio featurizer and add shift augmentor for DS2. #114
Conversation
1. Improve audio featurizer. 2. Add shift augmentor. 3. Update default argument to be the current best seggestion. 4. Add checkpoints with pass id.
| @@ -67,6 +67,54 @@ def from_file(cls, file): | |||
| return cls(samples, sample_rate) | |||
|
|
|||
| @classmethod | |||
| def slice_from_file(cls, file, start=None, end=None): | |||
There was a problem hiding this comment.
@reviewers:
No different for slice_from_file and make_silence. Only re-order them.
deep_speech_2/data_utils/audio.py
Outdated
| :type shift_ms: float | ||
| :raises ValueError: If shift_ms is longer than audio duration. | ||
| """ | ||
| if shift_ms / 1000.0 > self.duration: |
| extracting spectrogram features. | ||
| :type target_sample_rate: float | ||
| :param use_dB_normalization: Whether to normalize the audio to a certain | ||
| decibels before extracting the features. |
There was a problem hiding this comment.
Better to change decibels to dB for consistency
There was a problem hiding this comment.
For comments, full name decibels is used for clarity, while in arguments a short name of dB is used instead.
I think it makes sense?
| if audio_segment.sample_rate != self._target_sample_rate: | ||
| raise ValueError("Audio sample rate is not supported. " | ||
| "Turn allow_downsampling or allow up_sampling on.") | ||
| # decibel normalization |
There was a problem hiding this comment.
For comments, full name decibels is used for clarity, while in arguments a short name of dB is used instead.
I think it makes sense?
| :param use_dB_normalization: Whether to normalize the audio to a certain | ||
| decibels before extracting the features. | ||
| :type use_dB_normalization: bool | ||
| :param target_dB: Target audio decibels for normalization. |
There was a problem hiding this comment.
For comments, full name decibels is used for clarity, while in arguments a short name of dB is used instead.
I think it makes sense?
|
The experimental results for comparing old and new featurizer are as follows. We have a better convergence with the new featurizer.
|
Patchset for adding missing shift_perturb.py in PR #114.
resolve #113
Training experiment is in progress, and its results will be pasted here ASAP.