Following the general design of pytorch, our lib package can be extended easily to implement other new RNN models. Some models in this project are not complete. I'm still working on it. Note: This project bases on the project of GlenHGHUANG. It is a pytorch version of GlenHGHUANG's framework.
- Compiled Kaldi instance (instructions)
- Install Anaconda and create the environment with python 3.7, pytorch 1.3.
steps_sru/lib
: package for basic functionality of neural networks and the implementations of RNN modelssteps_sru/dataGenSequences_cxt_nolap_fbank.py
: data iteratersteps_sru/train_sru_nolap_12layer_1024_fbank.py
: acoustic modelingsteps_sru/decode_myseq_nolap_sru_dnn_fbank.sh
: decodersteps_sru/nnet-forward-myseq_nolap_sru_dnn.py
: HMM state posterior probability estimator
- Call the script from kaldi first: egs/dataset/s5/run.sh
- Call the script to generate Fbank features: steps/make_fbank.sh
- Call the script to perform speaker level MVN: steps/compute_cmvn_stats.sh
- Call the script: rnn_kt_sru_12layer.sh