Convert kaldi feature extraction and nnet3 models into Tensorflow Lite models. Currently aimed at converting kaldi's x-vector models and diarization pipelines to tensorflow models.
Apache License 2.0
20
stars
3
forks
source link
feat: build x-vector models from config file and initialize weights from nnet3 model files. #4
The kaldi models using TDNN, ReLU, BatchNorm and StatPooling layers can now be build from yml config files specifying the layers.
These models can then be intiialized with weights from kaldi's nnet3 model files.
Added two example config files for x-vector models available on the kaldi website.
Added script to convert a tensorflow SavedModel into a TF Lite model.
Added unit tests for the above.