Hugging Face's logo

Japanese ASR

non-profit
https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0

AI & ML interests

This repo contains models and datasets for Japanese ASR. See our main model https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0.

Organization Card
Community About org cards

Japanese ASR

This repository contains all the models and datasets for train/evaluate the Japanese ASR dataset generated through the process of achieving kotoba-whisper models. Following table shows CER comparison with different data size of ReazonSpeech used to distill openai/whisper-large-v3. The model names follows japanese-asr/distil-whisper-large-v3-ja-reazonspeech-{size of reazonspeech}.

CER

model CommonVoice 8 (Japanese test set) JSUT Basic 5000 ReazonSpeech (held out test set)
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-all 9.2 8.4 11.6
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-large 9.4 8.5 12.2
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-medium 10.9 11.3 14.8
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-small 30.2 39 40.7
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-tiny 94.8 96.3 96.7
openai/whisper-large-v3 8.5 7.1 14.9
openai/whisper-large-v2 9.7 8.2 28.1
openai/whisper-large 10 8.9 34.1
openai/whisper-medium 11.5 10 33.2
openai/whisper-base 28.6 24.9 70.4
openai/whisper-small 15.1 14.2 41.5
openai/whisper-tiny 53.7 36.5 137.9
reazon-research/reazonspeech-nemo-v2 9.1 7.4 11.2

WER

model CommonVoice 8 (Japanese test set) JSUT Basic 5000 ReazonSpeech (held out test set)
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-all 58.8 63.7 55.6
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-large 59.2 64.3 56.4
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-medium 64.6 72.1 63
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-small 85 94.2 82.1
japanese-asr/distil-whisper-large-v3-ja-reazonspeech-tiny 100 100 99
openai/whisper-large-v3 55.1 59.2 60.2
openai/whisper-large-v2 59.3 63.2 74.1
openai/whisper-large 61.1 66.4 74.9
openai/whisper-medium 63.4 69.5 76
openai/whisper-base 87.2 93 91.8
openai/whisper-small 74.2 81.9 83
openai/whisper-tiny 93.8 97.6 94.9
reazon-research/reazonspeech-nemo-v2 57.5 60.6 47.5

Note that kotoba-tech/kotoba-whisper-v1.0 is an alias of japanese-asr/distil-whisper-large-v3-ja-reazonspeech-large and kotoba-tech/kotoba-whisper-v2.0 is an alias of japanese-asr/distil-whisper-large-v3-ja-reazonspeech-all.

Please find more detailed results at kotoba-whisper codebase.

Collections 7

Bilingual (JA/EN) Bidirection ASR & Translation Dataset
These are the collection of the Bilingual ASR datasets labelled by the whisper-large-v3. The dataset consists of ASR and S2T translation tasks.

models 8