americasnlp22-asr-gn

Usage

import librosa
from espnet2.bin.asr_inference import Speech2Text

speech2text = Speech2Text.from_pretrained(model_tag="espnet/americasnlp22-asr-gn")
# 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

espnet/americasnlp22-asr-gn

This model was trained by Pavel Denisov using americasnlp22 recipe in espnet.

Demo: How to use in ESPnet2

Follow the ESPnet installation instructions if you haven't done that already.

cd espnet
git checkout fc62b1ce3e50c5ef8a2ac8cedb0d92ac41df54ca
pip install -e .
cd egs2/americasnlp22/asr1
./run.sh \
    --skip_data_prep false \
    --skip_train true \
    --download_model espnet/americasnlp22-asr-gn \
    --lang gn \
    --local_data_opts "--lang gn" \
    --train_set train_gn \
    --valid_set dev_gn \
    --test_sets dev_gn \
    --gpu_inference false \
    --inference_nj 8 \
    --lm_train_text data/train_gn/text \
    --bpe_train_text data/train_gn/text

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

RESULTS

Environments

• date: Sun Jun 5 12:17:58 CEST 2022
• python version: 3.9.13 (main, May 18 2022, 00:00:00) [GCC 11.3.1 20220421 (Red Hat 11.3.1-2)]
• espnet version: espnet 202204
• pytorch version: pytorch 1.11.0+cu115
• Git hash: d55704daa36d3dd2ca24ae3162ac40d81957208c
  • Commit date: Wed Jun 1 02:33:09 2022 +0200

asr_train_asr_transformer_raw_gn_bpe100_sp

WER

CER

TER

ASR config

<details><summary>expand</summary>

config: conf/train_asr_transformer.yaml
print_config: false
log_level: INFO
dry_run: false
iterator_type: sequence
output_dir: exp/asr_train_asr_transformer_raw_gn_bpe100_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: 15
patience: null
val_scheduler_criterion:
- valid
- loss
early_stopping_criterion:
- valid
- loss
- min
best_model_criterion:
-   - valid
    - cer_ctc
    - min
keep_nbest_models: 1
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.model.feature_extractor
- frontend.upstream.model.encoder.layers.0
- frontend.upstream.model.encoder.layers.1
- frontend.upstream.model.encoder.layers.2
- frontend.upstream.model.encoder.layers.3
- frontend.upstream.model.encoder.layers.4
- frontend.upstream.model.encoder.layers.5
- frontend.upstream.model.encoder.layers.6
- frontend.upstream.model.encoder.layers.7
- frontend.upstream.model.encoder.layers.8
- frontend.upstream.model.encoder.layers.9
- frontend.upstream.model.encoder.layers.10
- frontend.upstream.model.encoder.layers.11
- frontend.upstream.model.encoder.layers.12
- frontend.upstream.model.encoder.layers.13
- frontend.upstream.model.encoder.layers.14
- frontend.upstream.model.encoder.layers.15
- frontend.upstream.model.encoder.layers.16
- frontend.upstream.model.encoder.layers.17
- frontend.upstream.model.encoder.layers.18
- frontend.upstream.model.encoder.layers.19
- frontend.upstream.model.encoder.layers.20
- frontend.upstream.model.encoder.layers.21
num_iters_per_epoch: null
batch_size: 20
valid_batch_size: null
batch_bins: 200000
valid_batch_bins: null
train_shape_file:
- exp/asr_stats_raw_gn_bpe100_sp/train/speech_shape
- exp/asr_stats_raw_gn_bpe100_sp/train/text_shape.bpe
valid_shape_file:
- exp/asr_stats_raw_gn_bpe100_sp/valid/speech_shape
- exp/asr_stats_raw_gn_bpe100_sp/valid/text_shape.bpe
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_gn_sp/wav.scp
    - speech
    - sound
-   - dump/raw/train_gn_sp/text
    - text
    - text
valid_data_path_and_name_and_type:
-   - dump/raw/dev_gn/wav.scp
    - speech
    - sound
-   - dump/raw/dev_gn/text
    - text
    - text
allow_variable_data_keys: false
max_cache_size: 0.0
max_cache_fd: 32
valid_max_cache_size: null
optim: adamw
optim_conf:
    lr: 0.0001
scheduler: warmuplr
scheduler_conf:
    warmup_steps: 300
token_list:
- <blank>
- <unk>
- ▁
- a
- i
- e
- o
- ''''
- .
- u
- '"'
- p
- r
- n
- y
- h
- ▁"
- ▁o
- é
- re
- va
- pe
- s
- ra
- á
- he
- t
- mb
- g
- ka
- ã
- v
- ve
- je
- ▁ha
- te
- k
- ñ
- ha
- py
- ta
- ku
- ẽ
- ja
- pa
- O
- mi
- ó
- mo
- j
- ko
- ʼ
- ña
- me
- ma
- c
- M
- í
- H
- ú
- A
- ̃
- õ
- ý
- m
- P
- U
- ','
- ũ
- l
- ỹ
- N
- ĩ
- E
- I
- J
- L
- Á
- V
- S
- z
- '-'
- '?'
- Ñ
- R
- G
- Y
- T
- K
- C
- d
- “
- B
- ’
- ”
- D
- b
- f
- q
- <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
use_preprocessor: true
token_type: bpe
bpemodel: data/gn_token_list/bpe_unigram100/bpe.model
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: wav2vec2_url
        upstream_ckpt: https://dl.fbaipublicfiles.com/fairseq/wav2vec/xlsr2_300m.pt
    download_dir: ./hub
    multilayer_feature: true
    fs: 16k
specaug: null
specaug_conf: {}
normalize: utterance_mvn
normalize_conf: {}
model: espnet
model_conf:
    ctc_weight: 1.0
    lsm_weight: 0.0
    length_normalized_loss: false
    extract_feats_in_collect_stats: false
preencoder: linear
preencoder_conf:
    input_size: 1024
    output_size: 80
encoder: transformer
encoder_conf:
    input_layer: conv2d2
    num_blocks: 1
    linear_units: 2048
    dropout_rate: 0.2
    output_size: 256
    attention_heads: 8
    attention_dropout_rate: 0.2
postencoder: null
postencoder_conf: {}
decoder: rnn
decoder_conf: {}
required:
- output_dir
- token_list
version: '202204'
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}
}