The Kaldi Feature Written by Python
Please see the featurebin/
.
You should change the conf/
and write a .ini
file as a config file.
Note that you can write in several settings into one .ini
and you can use config_section
to
You should first prepare the wav.scp
for the project
which is similar with Kaldi
Note that the project do not support the pipe data for extracting features!
I will add this function in the following version
The project will write to some *.ark
files and a feats.scp
.
The features the project get will not be compressed,
however, you can use the feature_reader code to read the kaldi feature.
The
Kaldi
features compress is irreversible which means you cannot get the features before compressing, if you usewsj/steps/make_mfcc.sh
, etc., thecompress
parameter is always true unless you change it, there're no plan to add compress functions).
Please see the torch_feature/
and use the feature_kernels
.
The torch-api for computing the feature more quickly and
you can insert the feature-extractor into the model,
and you can even train them as I provide the requires_grad
!
-
DO NOT realize the
dither
for each, you can see the function but we will not use that, which may cause the great difficulty for debugging. -
DO NOT do the
vltn
part in our projects, so when we will get the single instance formel_banks
-
Use more value-call for the functions instead of reference-call, which is simpler for python programming. So you can see there're less
assert
to check the dim problems.
- Add pitches
- Merge more functions into one function, such as computer power spectrogram
- Add PLP features
- Use Subprocess to let the
wav
can be apipe data
- Add multiprocess part to improve the effectiveness
- More useful verbose and logging code
Tensorflow
support- Compare with
torchaudio
Please give me more suggestions and help improvement! You can email me and ask any questions in issue!
- Kaldi
- The kaldi-io code is from kaldi-python-io