akreal_swbd_da_hubert_conformer

Usage

import librosa
from espnet2.bin.asr_inference import Speech2Text

speech2text = Speech2Text.from_pretrained(model_tag="espnet/akreal_swbd_da_hubert_conformer")
# librosa resamples and mixes to one channel, so any file works; 16000 is
# what nearly every espnet recogniser is trained on - check this model's
# config if its audio is not 16 kHz
speech, rate = librosa.load("audio.wav", sr=16000, mono=True)
text, *_ = speech2text(speech)[0]
print(text)

ESPnet2 ASR model

akreal/espnet2_swbd_da_hubert_conformer

This model was trained by Pavel Denisov using swbd_da recipe in espnet.

Demo: How to use in ESPnet2

cd espnet
git checkout 08c6efbc6299c972301236625f9abafe087c9f9c
pip install -e .
cd egs2/swbd_da/asr1
./run.sh --skip_data_prep false --skip_train true --download_model espnet/akreal_swbd_da_hubert_conformer

<!-- Generated by scripts/utils/show_asr_result.sh -->

RESULTS

Environments

• date: Thu Jan 20 19:31:21 CET 2022
• python version: 3.8.12 (default, Aug 30 2021, 00:00:00) [GCC 11.2.1 20210728 (Red Hat 11.2.1-1)]
• espnet version: espnet 0.10.6a1
• pytorch version: pytorch 1.10.1+cu113
• Git hash: 08c6efbc6299c972301236625f9abafe087c9f9c
  • Commit date: Tue Jan 4 13:40:33 2022 +0100

asr_train_asr_raw_en_word_sp

WER

CER

TER

ASR config

<details><summary>expand</summary>

config: conf/tuning/train_asr_conformer_hubert_context3.yaml
print_config: false
log_level: INFO
dry_run: false
iterator_type: sequence
output_dir: exp/asr_train_asr_conformer_hubert_context3_raw_en_word_sp
ngpu: 1
seed: 0
num_workers: 1
num_att_plot: 3
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: 35
patience: null
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - valid
    - loss
    - min
keep_nbest_models: 7
nbest_averaging_interval: 0
grad_clip: 5.0
grad_clip_type: 2.0
grad_noise: false
accum_grad: 1
no_forward_run: false
resume: true
train_dtype: float32
use_amp: false
log_interval: null
use_matplotlib: true
use_tensorboard: true
use_wandb: false
wandb_project: null
wandb_id: null
wandb_entity: null
wandb_name: null
wandb_model_log_interval: -1
detect_anomaly: false
pretrain_path: null
init_param: []
ignore_init_mismatch: false
freeze_param:
- frontend.upstream
num_iters_per_epoch: null
batch_size: 20
valid_batch_size: null
batch_bins: 4000000
valid_batch_bins: null
train_shape_file:
- exp/asr_stats_context3_raw_en_word_sp/train/speech_shape
- exp/asr_stats_context3_raw_en_word_sp/train/text_shape.word
valid_shape_file:
- exp/asr_stats_context3_raw_en_word_sp/valid/speech_shape
- exp/asr_stats_context3_raw_en_word_sp/valid/text_shape.word
batch_type: numel
valid_batch_type: null
fold_length:
- 80000
- 150
sort_in_batch: descending
sort_batch: descending
multiple_iterator: false
chunk_length: 500
chunk_shift_ratio: 0.5
num_cache_chunks: 1024
train_data_path_and_name_and_type:
-   - dump/raw/train_context3_sp/wav.scp
    - speech
    - sound
-   - dump/raw/train_context3_sp/text
    - text
    - text
valid_data_path_and_name_and_type:
-   - dump/raw/valid_context3/wav.scp
    - speech
    - sound
-   - dump/raw/valid_context3/text
    - text
    - text
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
valid_max_cache_size: null
optim: adam
optim_conf:
    lr: 0.0001
scheduler: warmuplr
scheduler_conf:
    warmup_steps: 25000
token_list:
- <blank>
- <unk>
- statement
- backchannel
- opinion
- abandon
- agree
- yn_q
- apprec
- 'yes'
- uninterp
- close
- wh_q
- acknowledge
- 'no'
- yn_decl_q
- hedge
- backchannel_q
- sum
- quote
- affirm
- other
- directive
- repeat
- open_q
- completion
- rhet_q
- hold
- reject
- answer
- neg
- ans_dispref
- repeat_q
- open
- or
- commit
- maybe
- decl_q
- third_pty
- self_talk
- thank
- apology
- tag_q
- downplay
- <sos/eos>
init: null
input_size: null
ctc_conf:
    dropout_rate: 0.0
    ctc_type: builtin
    reduce: true
    ignore_nan_grad: true
joint_net_conf: null
model_conf:
    ctc_weight: 0.0
    extract_feats_in_collect_stats: false
use_preprocessor: true
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'
frontend: s3prl
frontend_conf:
    frontend_conf:
        upstream: hubert_large_ll60k
    download_dir: ./hub
    multilayer_feature: true
    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: utterance_mvn
normalize_conf: {}
preencoder: linear
preencoder_conf:
    input_size: 1024
    output_size: 80
encoder: conformer
encoder_conf:
    output_size: 512
    attention_heads: 8
    linear_units: 2048
    num_blocks: 12
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.1
    input_layer: conv2d
    normalize_before: true
    macaron_style: true
    pos_enc_layer_type: rel_pos
    selfattention_layer_type: rel_selfattn
    activation_type: swish
    use_cnn_module: true
    cnn_module_kernel: 31
postencoder: null
postencoder_conf: {}
decoder: transformer
decoder_conf:
    attention_heads: 8
    linear_units: 2048
    num_blocks: 6
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    self_attention_dropout_rate: 0.1
    src_attention_dropout_rate: 0.1
required:
- output_dir
- token_list
version: 0.10.5a1
distributed: false

</details>

Citing ESPnet

@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:

@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}
}