Usage

import librosa
from espnet2.bin.asr_inference import Speech2Text

speech2text = Speech2Text.from_pretrained(model_tag="espnet/MInDS-14_es-ES")
# 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)

RESULTS

Environments

date: Mon Mar 14 22:28:37 UTC 2022
python version: 3.8.12 | packaged by conda-forge | (default, Jan 30 2022, 23:42:07) [GCC 9.4.0]
espnet version: espnet 0.10.7a1
pytorch version: pytorch 1.10.1
Git hash: d5322b2dc4844dce1d14268b6848607e2a3dee21
- Commit date: Mon Mar 14 20:21:16 2022 +0000

asr_train_asr_raw_word

WER

dataset	Snt	Wrd	Corr	Sub	Del	Ins	Err	S.Err
inference_asr_model_valid.acc.ave_5best/test	49	4134	64.6	23.5	11.8	16.4	51.8	98.0
inference_asr_model_valid.acc.ave_5best/valid	47	4178	66.8	20.2	13.0	19.2	52.5	100.0

CER

dataset	Snt	Wrd	Corr	Sub	Del	Ins	Err	S.Err
inference_asr_model_valid.acc.ave_5best/test	49	8690	73.2	18.0	8.8	12.9	39.7	98.0
inference_asr_model_valid.acc.ave_5best/valid	47	8751	74.3	15.7	10.0	15.6	41.3	100.0

TER

dataset	Snt	Wrd	Corr	Sub	Del	Ins	Err	S.Err