sdqdlgj / asr-example.pytorch

Basic acoustic model implemented in pytorch

Pytorch examples for basic ASR

• compute-posterior.py gives outputs(posteriors) of the pytorch model which feed the command latgen-faster-mapped in kaldi.
• decode.sh combines kaldi command and pytorch model, generating the recognized lattice and scoring the WER/CER.
• model.py implements some basic acoustic models(CNN/DNN/RNN/TDNN/ResNet……)
• data/dataset.py is a simple wrapper of training corpus for DNN/CNN.
• prepare_*.py are some scripts for data pre-processing.
• updating……

TIMIT(test set)

MODEL	PER
DNN(3X1024) + BN	25.3%
DNN(4X512) + BN	24.1%
DNN(4X512) + BN + Dropout	23.8%
DNN(4X1024) + BN + Dropout	23.7%
CNN(K10,P6) + DNN(2X512)	23.1%
CNN(K10,P6) + DNN(2X512) + BN + Dropout	22.7%
GRU(3X256) + BN + Dropout	23.3%
GRU(4X256) + BN + Dropout	22.9%
GRU(3X512) + BN + Dropout	22.7%

THCHS30(test set)

MODEL	WER
CNN(K10,P6,C128) + DNN(2X512) + BN + Dropout	25.44%
CNN(K8,P6,C256) + DNN(2X512) + BN + Dropout	24.82%
GRU(3X512,T=20) + BN + Dropout	26.64%
GRU(3X512,T=100) + BN + Dropout	24.66%
GRU(10X256,T=100) + BN + Dropout	24.39%
Res-GRU(10X256,T=100) + BN + Dropout	24.24%
DNN(3X1024) + BN + Dropout	23.54%