返回模型
说明文档
注意:此配方使用此 PR https://github.com/k2-fsa/icefall/pull/349 中的代码进行训练
基于 icefall 的 WenetSpeech 数据集预训练 Transducer-Stateless2 模型
该模型使用 icefall 中的脚本,基于最新版本的 k2,在 WenetSpeech 的 L 子集上进行训练。
训练流程
主要代码库如下,我们将随着版本更新同步更新训练和解码脚本。
k2: https://github.com/k2-fsa/k2 icefall: https://github.com/k2-fsa/icefall lhotse: https://github.com/lhotse-speech/lhotse
- 安装 k2 和 lhotse,k2 安装指南请参考 https://k2.readthedocs.io/en/latest/installation/index.html,lhotse 请参考 https://lhotse.readthedocs.io/en/latest/getting-started.html#installation。我认为使用最新版本即可。同时请安装 icefall 中列出的依赖项。
- 克隆 icefall(https://github.com/k2-fsa/icefall) 并切换到上述显示的提交。
git clone https://github.com/k2-fsa/icefall
cd icefall
- 准备数据。
cd egs/wenetspeech/ASR
bash ./prepare.sh
- 训练
export CUDA_VISIBLE_DEVICES=\"0,1,2,3,4,5,6,7\"
./pruned_transducer_stateless2/train.py \
--world-size 8 \
--num-epochs 15 \
--start-epoch 0 \
--exp-dir pruned_transducer_stateless2/exp \
--lang-dir data/lang_char \
--max-duration 180 \
--valid-interval 3000 \
--model-warm-step 3000 \
--save-every-n 8000 \
--training-subset L
评估结果
以下是 WenetSpeech(dev、test-net 和 test-meeting)上的解码结果(WER%),我们通过对第 9 到 10 轮的模型进行平均获得了此结果。
WER 如下:
| dev | test-net | test-meeting | 备注 | |
|---|---|---|---|---|
| 贪心搜索 | 7.80 | 8.75 | 13.49 | --epoch 10, --avg 2, --max-duration 100 |
| 改进的束搜索 (beam size 4) | 7.76 | 8.71 | 13.41 | --epoch 10, --avg 2, --max-duration 100 |
| 快速束搜索 (1best) | 7.94 | 8.74 | 13.80 | --epoch 10, --avg 2, --max-duration 1500 |
| 快速束搜索 | 9.82 | 10.98 | 16.37 | --epoch 10, --avg 2, --max-duration 600 |
| 快速束搜索 | 6.88 | 7.18 | 11.77 | --epoch 10, --avg 2, --max-duration 600 |
| 快速束搜索 | 14.94 | 16.14 | 22.93 | --epoch 10, --avg 2, --max-duration 600 |
luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless2
作者 luomingshuang
↓ 0
♥ 4
创建时间: 2022-05-19 14:32:27+00:00
更新时间: 2022-10-13 06:43:37+00:00
在 Hugging Face 上查看文件 (145)
.gitattributes
README.md
data/lang_char/LG.pt
data/lang_char/Linv.pt
data/lang_char/tokens.txt
data/lang_char/words.txt
exp/cpu_jit_epoch_10_avg_2_torch_1.11.0.pt
exp/cpu_jit_epoch_10_avg_2_torch_1.7.1.pt
exp/decoder-epoch-10-avg-2.onnx
ONNX
exp/encoder-epoch-10-avg-2.onnx
ONNX
exp/export.sh
exp/joiner-epoch-10-avg-2.onnx
ONNX
exp/joiner_decoder_proj-epoch-10-avg-2.onnx
ONNX
exp/joiner_encoder_proj-epoch-10-avg-2.onnx
ONNX
exp/pretrained_epoch_10_avg_2.pt
exp_trained_with_M/pretrained_epoch_29_avg_11.pt
exp_trained_with_S/pretrained_epoch_29_avg_24.pt
log/fast_beam_search/errs-DEV-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/errs-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/errs-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/log-decode-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8-2022-05-19-23-31-13
log/fast_beam_search/recogs-DEV-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/recogs-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/recogs-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/wer-summary-DEV-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/wer-summary-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search/wer-summary-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-10-avg-2-beam-4-max-contexts-4-max-states-8.txt
log/fast_beam_search_nbest/errs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/errs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/errs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/log-decode-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-2022-07-18-16-24-35
log/fast_beam_search_nbest/recogs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/recogs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/recogs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/wer-summary-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/wer-summary-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest/wer-summary-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_LG/errs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/errs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/errs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/log-decode-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35-2022-07-26-11-26-29
log/fast_beam_search_nbest_LG/recogs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/recogs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/recogs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/wer-summary-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/wer-summary-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_LG/wer-summary-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-ngram-lm-scale-0.35.txt
log/fast_beam_search_nbest_oracle/errs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/errs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/errs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/log-decode-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64-2022-07-18-19-25-34
log/fast_beam_search_nbest_oracle/recogs-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/recogs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/recogs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/wer-summary-DEV-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/wer-summary-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/fast_beam_search_nbest_oracle/wer-summary-TEST_NET-beam_size_4-epoch-10-avg-2-beam-20.0-max-contexts-8-max-states-64.txt
log/greedy_search/errs-DEV-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/errs-TEST_MEETING-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/errs-TEST_NET-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/log-decode-epoch-10-avg-2-context-2-max-sym-per-frame-1-2022-05-19-22-45-39
log/greedy_search/recogs-DEV-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/recogs-TEST_MEETING-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/recogs-TEST_NET-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/wer-summary-DEV-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/wer-summary-TEST_MEETING-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/greedy_search/wer-summary-TEST_NET-greedy_search-epoch-10-avg-2-context-2-max-sym-per-frame-1.txt
log/modified_beam_search/errs-DEV-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/errs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/errs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/log-decode-epoch-10-avg-2-beam-4-2022-05-19-22-52-00
log/modified_beam_search/recogs-DEV-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/recogs-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/recogs-TEST_NET-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/wer-summary-DEV-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/wer-summary-TEST_MEETING-beam_size_4-epoch-10-avg-2-beam-4.txt
log/modified_beam_search/wer-summary-TEST_NET-beam_size_4-epoch-10-avg-2-beam-4.txt
log_trained_with_M/fast_beam_search/errs-DEV-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/errs-TEST_MEETING-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/errs-TEST_NET-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/log-decode-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8-2022-05-05-11-46-33
log_trained_with_M/fast_beam_search/log-decode-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8-2022-05-05-11-50-25
log_trained_with_M/fast_beam_search/recogs-DEV-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/recogs-TEST_MEETING-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/recogs-TEST_NET-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/wer-summary-DEV-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/wer-summary-TEST_MEETING-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/fast_beam_search/wer-summary-TEST_NET-beam_4.0_max_contexts_4_max_states_8-epoch-29-avg-11-beam-4.0-max-contexts-4-max-states-8.txt
log_trained_with_M/greedy_search/errs-DEV-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/errs-TEST_MEETING-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/errs-TEST_NET-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/log-decode-epoch-29-avg-11-context-2-max-sym-per-frame-1-2022-05-29-19-46-56
log_trained_with_M/greedy_search/recogs-DEV-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/recogs-TEST_MEETING-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/recogs-TEST_NET-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/wer-summary-DEV-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/wer-summary-TEST_MEETING-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/greedy_search/wer-summary-TEST_NET-greedy_search-epoch-29-avg-11-context-2-max-sym-per-frame-1.txt
log_trained_with_M/modified_beam_search/errs-DEV-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/errs-TEST_MEETING-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/errs-TEST_NET-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/log-decode-epoch-29-avg-11-beam-4-2022-05-05-10-23-19
log_trained_with_M/modified_beam_search/recogs-DEV-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/recogs-TEST_MEETING-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/recogs-TEST_NET-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/wer-summary-DEV-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/wer-summary-TEST_MEETING-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_M/modified_beam_search/wer-summary-TEST_NET-beam_size_4-epoch-29-avg-11-beam-4.txt
log_trained_with_S/fast_beam_search/errs-DEV-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/errs-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/errs-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/log-decode-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8-2022-05-29-21-10-39
log_trained_with_S/fast_beam_search/recogs-DEV-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/recogs-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/recogs-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/wer-summary-DEV-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/wer-summary-TEST_MEETING-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/fast_beam_search/wer-summary-TEST_NET-beam_4_max_contexts_4_max_states_8-epoch-29-avg-24-beam-4-max-contexts-4-max-states-8.txt
log_trained_with_S/greedy_search/errs-DEV-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/errs-TEST_MEETING-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/errs-TEST_NET-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/log-decode-epoch-29-avg-24-context-2-max-sym-per-frame-1-2022-05-29-20-15-37
log_trained_with_S/greedy_search/recogs-DEV-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/recogs-TEST_MEETING-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/recogs-TEST_NET-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/wer-summary-DEV-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/wer-summary-TEST_MEETING-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/greedy_search/wer-summary-TEST_NET-greedy_search-epoch-29-avg-24-context-2-max-sym-per-frame-1.txt
log_trained_with_S/modified_beam_search/errs-DEV-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/errs-TEST_MEETING-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/errs-TEST_NET-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/log-decode-epoch-29-avg-24-beam-4-2022-05-29-20-31-17
log_trained_with_S/modified_beam_search/recogs-DEV-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/recogs-TEST_MEETING-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/recogs-TEST_NET-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/wer-summary-DEV-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/wer-summary-TEST_MEETING-beam_size_4-epoch-29-avg-24-beam-4.txt
log_trained_with_S/modified_beam_search/wer-summary-TEST_NET-beam_size_4-epoch-29-avg-24-beam-4.txt
test_wavs/DEV_T0000000000.opus
test_wavs/DEV_T0000000000.wav
test_wavs/DEV_T0000000001.opus
test_wavs/DEV_T0000000001.wav
test_wavs/DEV_T0000000002.opus
test_wavs/DEV_T0000000002.wav
test_wavs/RESULTS.md