whisper-small.ja

Model Summary

whisper-small.ja is a Japanese Automatic Speech Recognition (ASR) model fine-tuned from OpenAI Whisper small.

Task: Automatic Speech Recognition (ASR)
Language: Japanese (JA)
Base Model: openai/whisper-small
Training Method: Full-parameter supervised fine-tuning

This model is trained on the Japanese subset of a publicly available multilingual speech dataset and aims to improve transcription performance for Japanese audio compared to the base whisper-small model.

⚠️ NSFW / Adult Content Notice This model's training data may contain adult (NSFW) speech content, and the model output may produce text not suitable for minors or general workplace use. Please verify your application scenario, audience, and compliance requirements before use.

Intended Use:

Transcription of Japanese speech that may contain adult content
Adult content review, organization, subtitling, or transcription (in compliant environments only)
Research purposes (e.g., ASR robustness with NSFW data)

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 solely responsible for ensuring that the acquisition, processing, and distribution of NSFW content comply with applicable laws.

How to Use

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch

model_id = "enet45/whisper-small.ja"

processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)

inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
    predicted_ids = model.generate(**inputs, language="ja")

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)

Recommended Inference Parameters:

language="ja"
task="transcribe"
beam_size=5

Training Data

Dataset: enet45/whisper-ja-en-zh-asr-translation
Subset: Japanese (JA) samples
Data Size:
- Total of approx. 39.5 hours of Japanese speech
- 60,499 valid Japanese samples
Average Audio Length: approx. 2.35 seconds
Data Type: mixture of read speech and conversational audio

Privacy and Copyright:

The dataset is publicly available
No additional identification or annotation of personal sensitive information
No extra manual filtering beyond standard preprocessing

Training Details

Base Model: openai/whisper-small
Training Strategy: Full-parameter fine-tuning (no LoRA / Adapter)
Epochs: 5
Learning Rate: 2e-6 (linear decay, 500 warmup steps)
Effective Batch Size: 32 (per-device batch size 8 × gradient accumulation 4)
Max Label Length: 64 tokens
Optimizer: AdamW (Transformers Trainer default)
Mixed Precision: FP16
Hardware Environment: single GPU (specific model depends on training setup)

Evaluation

Evaluation Method:
- Random 5% split of the dataset as validation
Metrics:
- Character Error Rate (CER)
- Word Error Rate (WER)

Validation Results (Best Model):

Metric	Value
CER ↓	1.60
WER ↓	1.27

Note: Results are based on internal validation and for model selection and comparison only.

Limitations

Performance may degrade on strong background noise, heavy accents, or unconventional pronunciation
Not optimized for specific professional domains (e.g., medical, legal, technical terms)
No guarantees for real-time streaming or very long audio sequences

Ethical Considerations

Training data may contain biases in language usage scenarios
Model outputs should not be considered authoritative or official transcriptions
Users must assess and comply with local laws regarding speech data and privacy

License

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

Citation

If you use this model in research or products, please cite the model and dataset sources.

Content Warning

This model may transcribe explicit sexual language or adult-oriented speech present in the input audio. It does not perform moderation, filtering, or content suppression. Users are solely responsible for ensuring compliant and lawful use.