icefall/RESULTS.md at a183d5bfd7f7c3ad16ed38c83a8a77933c085002

mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 18:12:19 +00:00

* add MMI to AIShell

* fix MMI decode graph

* export model

* typo

* fix code style

* typo

* fix data prepare to just use train text by uid

* use a faster way to get the intersection of train and aishell_transcript_v0.8.txt

* update AIShell result

* update

* typo

2021-12-04 14:43:04 +08:00

3.3 KiB

Raw Blame History

Results

Aishell training results (Conformer-MMI)

2021-12-04

(Pingfeng Luo): Result of https://github.com/k2-fsa/icefall/pull/140

The tensorboard log for training is available at https://tensorboard.dev/experiment/PSRYVbptRGynqpPRSykp1g

And pretrained model is available at https://huggingface.co/pfluo/icefall_aishell_mmi_model

The best decoding results (CER) are listed below, we got this results by averaging models from epoch 61 to 85, and using attention-decoder decoder with num_paths equals to 100.

	test
CER	4.94%

	lm_scale	attention_scale
test	1.1	0.3

You can use the following commands to reproduce our results:

git clone https://github.com/k2-fsa/icefall
cd icefall

cd egs/aishell/ASR
./prepare.sh

export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7,8"
python conformer_mmi/train.py --bucketing-sampler True \
                              --max-duration 200 \
                              --start-epoch 0 \
                              --num-epochs 90 \
                              --world-size 8

python conformer_mmi/decode.py --nbest-scale 0.5 \
                               --epoch 85 \
                               --avg 25 \
                               --method attention-decoder \
                               --max-duration 20 \
                               --num-paths 100

Aishell training results (Conformer-CTC)

2021-11-16

(Wei Kang): Result of https://github.com/k2-fsa/icefall/pull/30

Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc

The best decoding results (CER) are listed below, we got this results by averaging models from epoch 60 to 84, and using attention-decoder decoder with num_paths equals to 100.

	test
CER	4.26%

To get more unique paths, we scaled the lattice.scores with 0.5 (see https://github.com/k2-fsa/icefall/pull/10#discussion_r690951662 for more details), we searched the lm_score_scale and attention_score_scale for best results, the scales that produced the CER above are also listed below.

	lm_scale	attention_scale
test	0.3	0.9

You can use the following commands to reproduce our results:

git clone https://github.com/k2-fsa/icefall
cd icefall

cd egs/aishell/ASR
./prepare.sh

export CUDA_VISIBLE_DEVICES="0,1,2,3"
python conformer_ctc/train.py --bucketing-sampler True \
                              --max-duration 200 \
                              --start-epoch 0 \
                              --num-epochs 90 \
                              --world-size 4

python conformer_ctc/decode.py --nbest-scale 0.5 \
                               --epoch 84 \
                               --avg 25 \
                               --method attention-decoder \
                               --max-duration 20 \
                               --num-paths 100

Aishell training results (Tdnn-Lstm)

2021-09-13

(Wei Kang): Result of phone based Tdnn-Lstm model, https://github.com/k2-fsa/icefall/pull/30

Pretrained model is available at https://huggingface.co/pkufool/icefall_asr_aishell_conformer_ctc_lstm_ctc

The best decoding results (CER) are listed below, we got this results by averaging models from epoch 19 to 8, and using 1best decoding method.

	test
CER	10.16%

3.3 KiB Raw Blame History