CVTE_chain_model_finetune

• finetune chain model base on cvte open source model using aishell1 data
• using the open source GMM for frame align

Explanation

• cvte supply a chain model trained using more than 2000h audio data
• cvte supply a 3-gram LM model trained with 1000 GB text;
• this project does not need training any GMM series model
• this project support online cmvn, since "apply-cmvn-online" is used during the training and decoding

Install

• kaldi

Usage

• bash run_train_ft.sh to prepare the data、train LM(optinal,you can also use the cvte open source HCLG) and finetune chain model
• bash run_test.sh to decode the test data using the finetune model with apply-cmvn-online

Workflow

• prepare the lexicon if train LM with your own data(optional,you can also use the cvte open source HCLG)
• prepare the train data
• prepare phone sets, questions, L compilation(optinal,you can also use the cvte open source HCLG)
• train LM using kaldi_lm or SRILM(optinal,you can also use the cvte open source HCLG)
• make graph(optinal,you can also use the cvte open source HCLG)
• extract the 40 fbank
• generate speed-perturbed data (for alignment) and use it to align through cvte open source GMM(tri6b)
• get the alignments as lattices
• generate volume-perturbed data (for train)
• prepare config file for finetune such as model、den.fst、phone_lm.fst and tree
• funtune the chain model by some epoch with a little bit small learning rate
• decode the test data use finetune model with apply-cmvn-online

Result(aishell1 test data)

• cer 11.10%

Reference

• you can cd exp/chain/tdnn_ft/decode_test/scoreing_kaldi to view recognition results and decode cer
• also supply the finetune model,you can download it in the link 链接:https://pan.baidu.com/s/1NvsjV3R7PQy7q7YNITZS8g 密码:su1h
• cvte open source model link http://kaldi-asr.org/models/0002_cvte_chain_model.tar.gz