--- tags: - espnet - audio - automatic-speech-recognition language: jp datasets: - cejc_alt license: cc-by-4.0 --- ## ESPnet2 ASR model ### `fujie/espnet_asr_cbs_transducer_120303_hop132_cc0105` This model was trained by Shinya Fujie using cejc_alt recipe in [espnet](https://github.com/espnet/espnet/). ### Demo: How to use in ESPnet2 Follow the [ESPnet installation instructions](https://espnet.github.io/espnet/installation.html) if you haven't done that already. ```bash cd espnet git checkout 4c1c38f2c9c6a105ff4cffa8c833b0eb47f501a4 pip install -e . cd egs2/cejc_alt/asr1 ./run.sh --skip_data_prep false --skip_train true --download_model fujie/espnet_asr_cbs_transducer_120303_hop132_cc0105 ``` # RESULTS ## Environments - date: `Sun Mar 10 16:16:24 JST 2024` - python version: `3.11.7 (main, Dec 15 2023, 18:12:31) [GCC 11.2.0]` - espnet version: `espnet 202402` - pytorch version: `pytorch 2.1.0+cu121` - Git hash: `bf3653d6bd16c10a1df83f1db07e681374453f75` - Commit date: `Wed Mar 6 17:25:02 2024 +0900` ## exp/asr_train_asr_cbs_transducer_120303_hop132_cc0105 ### WER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval10f|953|11908|89.2|5.7|5.1|3.0|13.8|58.0| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval10m|957|16092|93.8|2.9|3.3|2.1|8.3|55.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval1_csj|1400|63362|94.9|3.0|2.1|1.2|6.3|69.5| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval20f|1466|18326|90.5|5.1|4.4|2.5|12.0|55.0| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval20m|1772|23756|89.0|5.8|5.2|2.8|13.8|56.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval2_csj|1413|64151|96.2|2.3|1.5|0.9|4.7|67.9| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval30f|1734|24116|93.6|3.4|3.0|2.3|8.8|48.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval30m|1688|20116|85.2|8.0|6.8|3.5|18.3|59.4| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval3_csj|1437|40131|96.3|2.0|1.8|1.2|4.9|52.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval40f|1477|20717|90.3|4.2|5.4|2.5|12.2|53.2| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval40m|1498|24747|92.4|3.5|4.1|2.3|9.9|55.7| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval50f|1450|26584|95.4|2.0|2.6|1.8|6.4|49.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval50m|1499|22572|92.0|4.1|4.0|2.4|10.4|54.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval60f|1335|21810|92.6|3.5|3.9|2.5|9.8|54.9| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval60m|1621|24151|89.5|5.0|5.4|2.3|12.8|62.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval70f|906|9542|88.7|5.7|5.6|3.4|14.7|53.4| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval70m|894|12490|92.9|3.5|3.5|2.6|9.7|51.6| ### CER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval10f|953|24583|91.5|3.5|5.0|3.1|11.6|58.0| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval10m|957|33749|94.9|1.6|3.5|2.4|7.5|55.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval1_csj|1400|139085|96.0|1.5|2.5|1.4|5.4|69.5| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval20f|1466|37024|92.3|3.1|4.6|2.6|10.4|55.0| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval20m|1772|47838|91.4|3.6|5.1|2.8|11.4|56.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval2_csj|1413|140081|97.0|1.0|2.0|1.2|4.2|67.9| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval30f|1734|48968|94.6|2.1|3.3|2.7|8.0|48.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval30m|1688|41067|88.4|4.9|6.7|3.5|15.1|59.4| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval3_csj|1437|86583|96.8|0.8|2.3|1.5|4.7|52.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval40f|1477|42609|91.7|2.8|5.5|2.4|10.7|53.2| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval40m|1498|51748|93.2|2.1|4.7|2.5|9.3|55.7| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval50f|1450|54181|95.8|1.4|2.8|1.9|6.1|49.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval50m|1499|46031|93.4|2.6|4.0|2.4|9.0|54.6| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval60f|1335|45028|93.9|2.0|4.2|2.7|8.9|54.9| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval60m|1621|49442|91.4|3.0|5.6|2.5|11.1|62.1| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval70f|906|19386|90.7|3.7|5.6|3.6|12.9|53.4| |decode_cbs_transducer_asr_model_valid.cer_transducer.ave_10best/eval70m|894|26203|94.1|2.1|3.7|3.0|8.9|51.6| ### TER |dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err| |---|---|---|---|---|---|---|---|---| ## ASR config
expand ``` config: myconf/train_asr_cbs_transducer_120303_hop132_silver11.yaml print_config: false log_level: INFO drop_last_iter: false dry_run: false iterator_type: sequence valid_iterator_type: null output_dir: exp/asr_train_asr_cbs_transducer_120303_hop132_cc0105 ngpu: 1 seed: 0 num_workers: 1 num_att_plot: 0 dist_backend: nccl dist_init_method: env:// dist_world_size: null dist_rank: null local_rank: 0 dist_master_addr: null dist_master_port: null dist_launcher: null multiprocessing_distributed: false unused_parameters: false sharded_ddp: false cudnn_enabled: true cudnn_benchmark: false cudnn_deterministic: true collect_stats: false write_collected_feats: false max_epoch: 100 patience: null val_scheduler_criterion: - valid - loss early_stopping_criterion: - valid - loss - min best_model_criterion: - - valid - cer_transducer - min keep_nbest_models: 10 nbest_averaging_interval: 0 grad_clip: 5 grad_clip_type: 2.0 grad_noise: false accum_grad: 6 no_forward_run: false resume: true train_dtype: float32 use_amp: false log_interval: null use_matplotlib: true use_tensorboard: true create_graph_in_tensorboard: false use_wandb: true wandb_project: espnet_ninjal wandb_id: null wandb_entity: null wandb_name: cejc_cbs_td_120303_hop132_cc0105 wandb_model_log_interval: -1 detect_anomaly: false use_lora: false save_lora_only: true lora_conf: {} pretrain_path: null init_param: - ./exp/asr_train_asr_cbs_transducer_081616_hop132/valid.cer_transducer.ave_10best.pth ignore_init_mismatch: false freeze_param: [] num_iters_per_epoch: null batch_size: 20 valid_batch_size: null batch_bins: 2000000 valid_batch_bins: null train_shape_file: - exp/asr_stats_raw_jp_word_cc0105/train/speech_shape - exp/asr_stats_raw_jp_word_cc0105/train/text_shape.word valid_shape_file: - exp/asr_stats_raw_jp_word_cc0105/valid/speech_shape - exp/asr_stats_raw_jp_word_cc0105/valid/text_shape.word batch_type: numel valid_batch_type: null fold_length: - 80000 - 150 sort_in_batch: descending shuffle_within_batch: false sort_batch: descending multiple_iterator: false chunk_length: 500 chunk_shift_ratio: 0.5 num_cache_chunks: 1024 chunk_excluded_key_prefixes: [] chunk_default_fs: null train_data_path_and_name_and_type: - - dump/raw/train_nodup_cc_01_05/wav.scp - speech - sound - - dump/raw/train_nodup_cc_01_05/text - text - text valid_data_path_and_name_and_type: - - dump/raw/train_dev_cc/wav.scp - speech - sound - - dump/raw/train_dev_cc/text - text - text allow_variable_data_keys: false max_cache_size: 0.0 max_cache_fd: 32 allow_multi_rates: false valid_max_cache_size: null exclude_weight_decay: false exclude_weight_decay_conf: {} optim: adam optim_conf: lr: 0.002 scheduler: warmuplr scheduler_conf: warmup_steps: 25000 token_list: - - - - '|' - ー - ン - イ - ト - カ - ノ - - テ - デ - タ - シ - ス - ナ - ッ - コ - オ - ニ - マ - ワ - ガ - ク - モ - ー+F - ル - キ - レ - エ+F - ラ - リ - ア - ケ - ツ - ソ - ユ - ド - サ - セ - ヨ - ダ - エ - チ - ジ - ア+F - ノ+F - ネ - ホ - マ+F - ハ - ゴ - ミ - ロ - ブ - バ - ヤ - ヒ - メ - ウ - フ - ショ - ジョ - ジュ - ズ - ゲ - シュ - ム - チョ - ト+F - キョ - グ - パ - ベ - シャ - ゼ - ソ+F - ン+F - ギ - ザ - ビ - キュ - ボ - リョ - ヘ - ゾ - プ - ン+D - チュ - ジャ - ウ+F - オ+F - ッ+F - ヒョ - チャ - イ+D - ヌ - ス+D - ポ - ピ - ディ - ティ - ギョ - ニュ - オ+D - イ+F - ー+D - ヒャ - シ+D - ペ - ッ+D - ウ+D - ア+D - カ+D - キャ - ク+D - コ+D - ナ+D - ツ+D - エ+D - ト+D - ビョ - ジェ - リュ - タ+D - ピョ - ハ+D - ヒ+D - ファ - ノ+D - キ+D - ニ+D - ギャ - ハ+F - モ+D - フィ - ソ+D - フ+D - ワ+D - ホ+D - ジ+D - マ+D - ヨ+D - デ+D - サ+D - ガ+D - ユ+D - セ+D - フォ - ム+D - ダ+D - テ+D - チ+D - ヤ+D - ケ+D - トゥ - ル+D - ラ+D - ウォ - リャ - ミ+D - ド+D - シュ+D - リ+D - ズ+D - ヘ+F - ウェ - レ+D - ピュ - ブ+D - フェ - ミョ - グ+D - ヌ+D - トゥ+D - テュ - ヘ+D - ロ+D - チェ - ゴ+D - ジュ+D - ミュ - ビャ - ネ+F - ピャ - ショ+D - メ+D - ミャ - ギュ - ネ+D - バ+D - スィ - ゲ+D - ビュ - ニョ - ジョ+D - チョ+D - ス+F - ゼ+D - デ+F - キョ+D - ヤ+F - チュ+D - プ+D - ワ+F - ギ+D - ウィ - ベ+D - シェ - ボ+D - パ+D - ドゥ+D - ニャ - シャ+D - ドゥ - ザ+D - ヒョ+D - レ+F - ツォ - ビ+D - ド+F - ニュ+D - キュ+D - リョ+D - デュ - ヒュ - ディ+D - ゾ+D - ティ+D - フ+F - ラ+F - ナ+F - ピ+D - リュ+D - ヒャ+D - ジャ+D - ヒュ+D - チャ+D - ツァ - ポ+D - ニョ+D - ツェ - ヌ+F - ズィ - キャ+D - ホ+F - ペ+D - ヴィ - ツ+F - ギョ+D - ファ+D - ウェ+D - ウォ+D - ツォ+F - ジェ+D - メ+F - フィ+D - バ+F - ニャ+D - ギャ+D - ビョ+D - ツィ - フォ+D - スィ+D - ウィ+D - リャ+D - モ+F - チェ+D - フュ - テュ+D - ロ+F - デュ+D - シェ+D - イェ - ム+F - ニェ - ツォ+D - トゥ+F - カ+F - ミャ+D - ミョ+D - ギュ+D - ミュ+D - ツァ+D - フェ+D - ガ+F - クヮ - ヨ+F - テ+F - ヒ+F - ズィ+D - グヮ - ウェ+F - ビュ+D - イェ+D - ユ+F - イェ+F - ツェ+D - パ+F - ヴァ - チョ+F - ニョ+F - ダ+F - ニェ+D - ル+F - ゼ+F - ゾ+F - ニェ+F - リャ+F - ミャ+F - ヴェ - ショ+F - キャ+F - ゲ+F - ピュ+D - ク+F - ニャ+F - ケ+F - ヴ - チャ+F - タ+F - グ+F - ヴォ - ミェ - ヒャ+F - ファ+F - フェ+F - ビャ+D - ブ+F - ズ+F - ジェ+F - ピャ+D - ツィ+D - リ+F - セ+F - サ+F - ドゥ+F - ウォ+F - グヮ+D - ベ+F - ザ+F - クヮ+D - ヒェ+D - シ+F - フュ+D - ヴィ+D - テュ+F - ミェ+D - ボ+F - ジャ+F - ヴァ+D - ジ+F - チ+F - ゴ+F - ピョ+D - ヒェ - ニ+F - シュ+F - ミュ+F - init: null input_size: null ctc_conf: dropout_rate: 0.0 ctc_type: builtin reduce: true ignore_nan_grad: null zero_infinity: true brctc_risk_strategy: exp brctc_group_strategy: end brctc_risk_factor: 0.0 joint_net_conf: joint_space_size: 640 use_preprocessor: true use_lang_prompt: false use_nlp_prompt: false token_type: word bpemodel: null non_linguistic_symbols: null cleaner: null g2p: null speech_volume_normalize: null rir_scp: null rir_apply_prob: 1.0 noise_scp: null noise_apply_prob: 1.0 noise_db_range: '13_15' short_noise_thres: 0.5 aux_ctc_tasks: [] frontend: default frontend_conf: hop_length: 132 fs: 16k specaug: specaug specaug_conf: apply_time_warp: true time_warp_window: 5 time_warp_mode: bicubic apply_freq_mask: true freq_mask_width_range: - 0 - 30 num_freq_mask: 2 apply_time_mask: true time_mask_width_range: - 0 - 40 num_time_mask: 2 normalize: global_mvn normalize_conf: stats_file: exp/asr_stats_raw_jp_word_cc0105/train/feats_stats.npz model: espnet model_conf: ctc_weight: 0.0 report_cer: true report_wer: true preencoder: null preencoder_conf: {} encoder: contextual_block_conformer encoder_conf: output_size: 256 attention_heads: 4 linear_units: 2048 num_blocks: 12 dropout_rate: 0.1 positional_dropout_rate: 0.1 attention_dropout_rate: 0.0 input_layer: conv2d normalize_before: true activation_type: swish macaron_style: true use_cnn_module: true cnn_module_kernel: 15 block_size: 18 hop_size: 3 look_ahead: 3 init_average: true ctx_pos_enc: true postencoder: null postencoder_conf: {} decoder: transducer decoder_conf: rnn_type: lstm num_layers: 1 hidden_size: 512 dropout: 0.1 dropout_embed: 0.2 preprocessor: default preprocessor_conf: {} required: - output_dir - token_list version: '202402' distributed: false ```
### Citing ESPnet ```BibTex @inproceedings{watanabe2018espnet, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, title={{ESPnet}: End-to-End Speech Processing Toolkit}, year={2018}, booktitle={Proceedings of Interspeech}, pages={2207--2211}, doi={10.21437/Interspeech.2018-1456}, url={http://dx.doi.org/10.21437/Interspeech.2018-1456} } ``` or arXiv: ```bibtex @misc{watanabe2018espnet, title={ESPnet: End-to-End Speech Processing Toolkit}, author={Shinji Watanabe and Takaaki Hori and Shigeki Karita and Tomoki Hayashi and Jiro Nishitoba and Yuya Unno and Nelson Yalta and Jahn Heymann and Matthew Wiesner and Nanxin Chen and Adithya Renduchintala and Tsubasa Ochiai}, year={2018}, eprint={1804.00015}, archivePrefix={arXiv}, primaryClass={cs.CL} } ```