zr_jin 0f1bc6f8af
Multi_zh-Hans Recipe (#1238)
* Init commit for recipes trained on multiple zh datasets.

* fbank extraction for thchs30

* added support for aishell1

* added support for aishell-2

* fixes

* fixes

* fixes

* added support for stcmds and primewords

* fixes

* added support for magicdata

script for fbank computation not done yet

* added script for magicdata fbank computation

* file permission fixed

* updated for the wenetspeech recipe

* updated

* Update preprocess_kespeech.py

* updated

* updated

* updated

* updated

* file permission fixed

* updated paths

* fixes

* added support for kespeech dev/test set fbank computation

* fixes for file permission

* refined support for KeSpeech

* added scripts for BPE model training

* updated

* init commit for the multi_zh-cn zipformer recipe

* disable speed perturbation by default

* updated

* updated

* added necessary files for the zipformer recipe

* removed redundant wenetspeech M and S sets

* updates for multi dataset decoding

* refined

* formatting issues fixed

* updated

* minor fixes

* this commit finalize the recipe (hopefully)

* fixed formatting issues

* minor fixes

* updated

* using soft links to reduce redundancy

* minor updates

* using soft links to reduce redundancy

* minor updates

* minor updates

* using soft links to reduce redundancy

* minor updates

* Update README.md

* minor updates

* Update egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* minor updates

* minor fixes

* fixed a formatting issue

* Update preprocess_kespeech.py

* Update prepare.sh

* Update egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* Update egs/multi_zh-hans/ASR/local/preprocess_kespeech.py

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>

* removed redundant files

* symlinks added

* minor updates

* added CI tests for `multi_zh-hans`

* minor fixes

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

* Update run-multi-zh_hans-zipformer.sh

---------

Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
2023-09-13 11:57:05 +08:00

2.1 KiB

Results

Multi Chinese datasets char-based training results (Non-streaming) on zipformer model

This is the pull request #1238 in icefall.

Non-streaming

Best results (num of params : ~69M):

The training command:

./zipformer/train.py \
  --world-size 4 \
  --num-epochs 20 \
  --use-fp16 1 \
  --max-duration 600 \
  --num-workers 8

The decoding command:

./zipformer/decode.py \
  --epoch 20 \
  --avg 1

Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model ( # tokens is 2000, byte fallback enabled).

Datasets aidatatang _200zh aidatatang _200zh alimeeting alimeeting aishell-1 aishell-1 aishell-2 aishell-2 aishell-4 magicdata magicdata kespeech-asr kespeech-asr kespeech-asr WenetSpeech WenetSpeech WenetSpeech
Zipformer CER (%) dev test eval test dev test dev test test dev test dev phase1 dev phase2 test dev test meeting test net
3.2 3.67 23.15 24.78 2.91 3.04 3.59 4.03 15.68 3.68 3.12 6.69 3.19 8.01 9.32 7.05 8.78

The pre-trained model is available here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2