espnet/shihlun_asr_whisper_medium_finetuned_chime4

import librosa
from espnet2.bin.asr_inference import Speech2Text

speech2text = Speech2Text.from_pretrained(model_tag="espnet/shihlun_asr_whisper_medium_finetuned_chime4")
# 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)

cd espnet
pip install -e .
cd egs2/chime4/asr1

train_set=tr05_multi_noisy_si284 # tr05_multi_noisy (original training data) or tr05_multi_noisy_si284 (add si284 data)
valid_set=dt05_multi_isolated_1ch_track
test_sets="dt05_real_isolated_1ch_track dt05_simu_isolated_1ch_track et05_real_isolated_1ch_track et05_simu_isolated_1ch_track"

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 \
    --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}" "$@"