placebokkk / ctc-asr

pytorch CTC implementation for ASR. Use eesen's fst decoder framework

ASR E2E acoustic model training with CTC in pytorch

I am a newbie in pytorch. I create this repo to learn pythorch and E2E asr modeling.

For practical E2E ASR job, please refer to better tools such as espnet.

Current Feature

• Very simple pytorch implementation without optimization
• lstm+ctc, same config with eesen
• kaldi feature extraction (40 fbank+delta+cmvn)
• eesen TLG fst decoding

Usage

1. Download and install Eesen

2. in path.sh file, set EESEN_ROOT to eesen path.

export EESEN_ROOT='your/eesen/root/dir'

3. run run_ctc_phn.sh

TODO

• CTC beamdecoder, phone/char/word LM
• Use RNN-t to involve LM.
• E2E structure
• Try other Net structure.