japanese-asr/en-cascaded-s2t-translation

Cascaded English Speech2Text Translation

This is a pipeline for speech-to-text translation from English speech to any target language text based on the cascaded approach, that consists of ASR and translation. The pipeline employs distil-whisper/distil-large-v3 for ASR (English speech -> English text) and facebook/nllb-200-3.3B for text translation. The input must be English speech, while the translation can be in any languages NLLB trained on. Please find the all available languages and their language codes here.

Model for Japanese speech translation is available at ja-cascaded-s2t-translation.

Benchmark

The folloiwng table shows CER computed over the reference and predicted translation for translating English speech to Japanese text task (subsets of CoVoST2 and Fleurs) with different size of NLLB along with OpenAI Whisper models.

model	CoVoST2 (En->Ja)	Fleurs (En->JA)
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-3.3B)	62.4	63.5
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-1.3B)	64.4	67.2
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-distilled-1.3B)	62.4	62.9
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-distilled-600M)	63.4	66.2
openai/whisper-large-v3	178.9	209.5
openai/whisper-large-v2	179.6	201.8
openai/whisper-large	178.7	201.8
openai/whisper-medium	178.7	202
openai/whisper-small	178.9	206.8
openai/whisper-base	179.5	214.2
openai/whisper-tiny	185.2	200.5

See https://github.com/kotoba-tech/kotoba-whisper for the evaluation detail.

Inference Speed

Due to the nature of cascaded approach, the pipeline has additional complexity compared to the single end2end OpenAI whisper models for the sake of high accuracy. Following table shows the mean inference time in second averaged over 10 trials on audio sample with different durations.

model	10	30	60	300
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-3.3B)	0.173	0.247	0.352	1.772
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-1.3B)	0.173	0.24	0.348	1.515
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-distilled-1.3B)	0.17	0.245	0.348	1.882
japanese-asr/en-cascaded-s2t-translation (facebook/nllb-200-distilled-600M)	0.108	0.179	0.283	1.33
openai/whisper-large-v3	0.061	0.184	0.372	1.804
openai/whisper-large-v2	0.062	0.199	0.415	1.854
openai/whisper-large	0.062	0.183	0.363	1.899
openai/whisper-medium	0.045	0.132	0.266	1.368
openai/whisper-small	0.135	0.376	0.631	3.495
openai/whisper-base	0.054	0.108	0.231	1.019
openai/whisper-tiny	0.045	0.124	0.208	0.838

Usage

Here is an example to translate English speech into Japanese text translation. First, download a sample speech.

wget https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval/resolve/main/sample.wav -O sample_en.wav

Then, run the pipeline as below.

from transformers import pipeline

# load model
pipe = pipeline(
    model="japanese-asr/en-cascaded-s2t-translation",
    model_translation="facebook/nllb-200-distilled-600M",
    tgt_lang="jpn_Jpan",
    model_kwargs={"attn_implementation": "sdpa"},
    chunk_length_s=15,
    trust_remote_code=True,
)

# translate
output = pipe("./sample.wav")

Other NLLB models can be used by setting model_translation such as following.

• facebook/nllb-200-3.3B
• facebook/nllb-200-distilled-600M
• facebook/nllb-200-distilled-1.3B
• facebook/nllb-200-1.3B