ESPnet2 ASR model

espnet/shihlun_asr_whisper_medium_finetuned_librispeech100

This model was trained by Shih-Lun Wu (slseanwu) using the librispeech_100 recipe in espnet.

Demo: How to use in ESPnet2

cd espnet
pip install -e .
cd egs2/librispeech_100/asr1

train_set="train_clean_100"
valid_set="dev"
test_sets="test_clean test_other dev_clean dev_other"
asr_tag=whisper_medium_finetune_lr1e-5_adamw_wd1e-2_3epochs
asr_config=conf/tuning/train_asr_whisper_full.yaml
inference_config=conf/decode_asr_whisper_noctc_greedy.yaml

./asr.sh \
    --skip_data_prep false \
    --skip_train true \
    --skip_eval false \
    --lang en \
    --ngpu 1 \
    --nj 4 \
    --stage 1 \
    --stop_stage 13 \
    --gpu_inference true \
    --inference_nj 1 \
    --token_type whisper_multilingual \
    --feats_normalize '' \
    --max_wav_duration 30 \
    --speed_perturb_factors "0.9 1.0 1.1" \
    --audio_format "flac.ark" \
    --feats_type raw \
    --use_lm false \
    --cleaner whisper_en \
    --asr_tag "${asr_tag}" \
    --asr_config "${asr_config}" \
    --inference_config "${inference_config}" \
    --inference_asr_model valid.acc.ave.pth \
    --train_set "${train_set}" \
    --valid_set "${valid_set}" \
    --test_sets "${test_sets}" "$@"

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

RESULTS

Environments

• date: Mon Jan 9 23:06:34 CST 2023
• python version: 3.9.13 (main, Aug 25 2022, 23:26:10) [GCC 11.2.0]
• espnet version: espnet 202211
• pytorch version: pytorch 1.12.1
• Git hash: d89be931dcc8f61437ac49cbe39a773f2054c50c
  • Commit date: Mon Jan 9 11:06:45 2023 -0600

asr_whisper_medium_finetune_lr1e-5_adamw_wd1e-2_3epochs

WER

CER