wav2vec2-ja-cv25-reazon-mixed

wav2vec2-ja-cv25-reazon-mixed

Japanese ASR model based on facebook/wav2vec2-large-xlsr-53.

• Task: Automatic Speech Recognition (CTC)
• Language: Japanese
• License: apache-2.0
• Base model: facebook/wav2vec2-large-xlsr-53
• Dataset:
  • Common Voice Scripted Speech 25.0 Japanese
  • ReazonSpeech(small)
• Decoding: greedy CTC, without external language model

Usage

import torch
import librosa
from transformers import AutoModelForCTC, Wav2Vec2Processor

repo_id = "takehika/wav2vec2-ja-cv25-reazon-mixed"

processor = Wav2Vec2Processor.from_pretrained(repo_id)
model = AutoModelForCTC.from_pretrained(repo_id)

speech, _ = librosa.load("sample.wav", sr=16_000)
inputs = processor(speech, sampling_rate=16_000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(**inputs).logits

pred_ids = torch.argmax(logits, dim=-1)
text = processor.batch_decode(pred_ids)[0]
print(text)

Data

• Dataset:
  • Common Voice Scripted Speech 25.0 Japanese
  • ReazonSpeech(small)

Text normalization:

• Unicode NFKC normalization
• Whitespace cleanup
• Keep Japanese characters, digits, and selected punctuation
• Remove most other symbols

Tokenizer/vocab behavior:

• Character-level CTC vocab
• Built from normalized training text
• Space is mapped to |
• Includes digits 0-9
• Includes 、, 。, 「, 」
• Does not include ？, ！ in the final saved vocab

Evaluation

External evaluation results saved in this project:

• Common Voice test: CER = 0.3343 on 8,957 samples
• JSUT basic5000: CER = 0.2632 on 4,999 samples

CER was computed using nearly the same text normalization as training preprocessing: both references and predictions were Unicode NFKC-normalized, whitespace-cleaned, | was mapped to a space, and only Japanese characters, digits 0-9, and selected punctuation were retained.

Transcription Examples

Common Voice test

Reference: 私はその人の記憶を呼び起すごとに、すぐ「先生」といいたくなる。
Prediction: 私はその人の記憶を呼び起こすごとにすぐ先生とい痛くなる。

Reference: 元気の出る曲をかけて
Prediction: 元金の出る曲をかけて

Reference: 日本語は美しいです。
Prediction: 日本語は美しいです。

Reference: 札幌スクーリングへ行く
Prediction: 札幌札幌ロスクーリグへ行く

Intended Use and Limitations

• Intended use: Japanese ASR for 16 kHz speech.
• Performance may degrade on noisy speech, conversational speech, dialect-heavy speech, or domains far from Common Voice Scripted Speech and ReazonSpeech.
• Text style: punctuation handling and formatting may differ from natural writing due to CTC-style decoding and preprocessing.
• Reliability: outputs may include transcription errors; human review is recommended for high-stakes use.

Attribution

• Base model:
  • facebook/wav2vec2-large-xlsr-53 - Apache-2.0
• Dataset:
  • Common Voice Scripted Speech 25.0 Japanese - CC0-1.0
  • ReazonSpeech - CDLA-Sharing-1.0 (see the dataset page for license/terms before use)

License

This model is licensed under Apache-2.0.