import librosa
import numpy as np
from transformers import AutoProcessor, Wav2Vec2ForCTC

model = Wav2Vec2ForCTC.from_pretrained(
    "reazon-research/japanese-wav2vec2-base-rs35kh",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
).to("cuda")
processor = AutoProcessor.from_pretrained("reazon-research/japanese-wav2vec2-base-rs35kh")

audio, _ = librosa.load(audio_filepath, sr=16_000)
audio = np.pad(audio, pad_width=int(0.5 * 16_000))  # Recommend to pad audio before inference
input_values = processor(
    audio,
    return_tensors="pt",
    sampling_rate=16_000
).input_values.to("cuda").to(torch.bfloat16)

with torch.inference_mode():
    logits = model(input_values).logits.cpu()
predicted_ids = torch.argmax(logits, dim=-1)[0]
transcription = processor.decode(predicted_ids, skip_special_tokens=True)

Test Results

We report the Character Error Rate (CER) of our model and the other wav2vec2 families.

Model	#Prameters⬇	AVERAGE⬇	JSUT-BASIC5000⬇	Common Voice⬇	TEDxJP-10K⬇
reazon-research/japanese-wav2vec2-base-rs35kh	96.7M	20.40%	13.22%	23.76%	24.23%
Ivydata/wav2vec2-large-xlsr-53-japanese	318M	24.23%	13.83%	18.15%	40.72%
jonatasgrosman/wav2vec2-large-xlsr-53-japanese	317M	31.82%	4.25%	40.58%	50.63%
vumichien/wav2vec2-large-xlsr-japanese	318M	39.87%	4.21%	53.29%	62.12%

We also report the CER for long-form speech.

Model	#Prameters⬇	JSUT-BOOK⬇
reazon-research/japanese-wav2vec2-base-rs35kh	96.7M	82.84%
Ivydata/wav2vec2-large-xlsr-53-japanese	318M	65.60%
jonatasgrosman/wav2vec2-large-xlsr-53-japanese	317M	46.20%
vumichien/wav2vec2-large-xlsr-japanese	318M	46.52%

Citation

bibtex

@misc{reazon-research-japanese-wav2vec2-base-rs35kh,
  title={japanese-wav2vec2-base-rs35kh},
  author={Sasaki, Yuta},
  url = {https://huggingface.co/reazon-research/japanese-wav2vec2-base-rs35kh},
  year = {2024}
}

License

Apaceh Licence 2.0