faster-whisper-small.ja

Model Summary

faster-whisper-small.ja is a faster inference version of the Japanese ASR model whisper-small.ja, converted using ct2-transformers-converter (float16).

• Task: Automatic Speech Recognition (ASR)
• Language: Japanese (JA)
• Base Model: enet45/whisper-small.ja
• Conversion Method: Converted for faster inference using ct2-transformers-converter with float16 quantization

This model provides lower latency inference while keeping transcription quality close to the original whisper-small.ja model.

⚠️ NSFW / Adult Content Notice
This model may output text containing adult (NSFW) content, inherited from the base model. Do not use in contexts where adult content is inappropriate.

Intended Use:

• Fast transcription of Japanese speech, including adult content
• Adult content review, organization, subtitling, or transcription (compliant environments only)
• Research and prototyping with faster ASR inference

Not Intended For:

• Products or services aimed at minors
• Public display, educational settings, or default user-generated content
• Any use violating local laws or regulations

Users are responsible for ensuring compliance with laws when handling NSFW content.

---

How to Use

from faster_whisper import WhisperModel

model_path = "/workspace/output-ct2"
model = WhisperModel(model_path, device="cuda", compute_type="float16")

# audio: path to a WAV/FLAC/MP3 file
segments, info = model.transcribe(audio, language="ja", beam_size=5)

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

Recommended Inference Parameters:

• language="ja"
• beam_size=5 This model is optimized for faster CPU/GPU inference compared to the original PyTorch version.

Training Data & Base Model

This model inherits training data and characteristics from whisper-small.ja:

• Dataset: enet45/whisper-ja-en-zh-asr-translation (Japanese subset)

• Training Method: Full-parameter fine-tuning on whisper-small

• Data Size: ~39.5 hours, 60,499 valid samples

• Average Audio Length: 2.35s

• Data Type: mixture of read speech and conversational audio

Privacy and Copyright:

• Publicly available datasets

• No additional sensitive information added

• No extra manual filtering beyond preprocessing

Evaluation

Evaluation metrics are inherited from the base model (whisper-small.ja) internal validation split:

Note: Faster inference conversion may slightly change timing or memory usage but does not significantly affect transcription quality.

Limitations

• Same limitations as the base model apply: background noise, strong accents, and unconventional pronunciations may reduce accuracy

• Not specifically optimized for professional domain terminology

• Conversion does not guarantee perfect parity with PyTorch version, minor differences possible

Ethical Considerations

• Model inherits NSFW content from base model

• Outputs should not be treated as authoritative

• Users are responsible for complying with local laws and ethical guidelines

License

This model inherits the license of the base model: Apache License 2.0.

Citation

If using this model, please cite both the original Whisper model and the whisper-small.ja fine-tuned dataset.

Content Warning

This model may transcribe explicit sexual language or adult-oriented speech. It does not perform content moderation. Users must ensure lawful and compliant use.