enet45/faster-kotoba-whisper-v2.1

This repository provides a CTranslate2 / faster-whisper conversion of:

• Base model: enet45/kotoba-whisper-v2.1

The model is optimized for high-performance inference using faster-whisper, enabling significantly lower latency and reduced memory usage compared to standard PyTorch inference.

⚠️ Content Warning: This model was trained on NSFW (Not Safe For Work) Japanese audio transcripts, including explicit adult content. It may produce sexually explicit or offensive transcriptions. Use only in environments where such output is acceptable and properly controlled.

---

Model Summary

• Architecture: Whisper (Encoder–Decoder)
• Framework: CTranslate2 (converted from Transformers)
• Parameters: ~756M
• Language: Japanese
• Task: Automatic Speech Recognition (ASR)
• Vocabulary size: 51,866
• Max input frames: 1500
• Max target length: 448
• Precision: Converted for FP16 inference (recommended)

Original training configuration:

• 32 encoder layers
• 2 decoder layers
• d_model: 1280
• 20 attention heads
• FFN dim: 5120
• Activation: GELU
• Gradient checkpointing enabled during training

---

Why faster-whisper?

Compared to standard 🤗 Transformers inference:

• ⚡ Lower latency
• 💾 Reduced memory footprint
• 🧠 Efficient batch inference
• 🖥️ Optimized for GPU and CPU
• 🔄 Supports INT8 / FP16 quantization

---

Installation

pip install faster-whisper

For GPU inference, ensure:

• CUDA ≥ 11.x
• Proper NVIDIA drivers installed

---

Usage

Basic Transcription

from faster_whisper import WhisperModel

model_size_or_path = "enet45/faster-kotoba-whisper-v2.1"

model = WhisperModel(
    model_size_or_path,
    device="cuda",          # "cpu" or "cuda"
    compute_type="float16"  # "float16", "int8", or "int8_float16"
)

segments, info = model.transcribe(
    "audio.wav",
    language="ja",
    beam_size=5
)

print("Detected language:", info.language)

for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

---

Recommended Settings

GPU (High Performance)

WhisperModel(..., device="cuda", compute_type="float16")

CPU (Memory Efficient)

WhisperModel(..., device="cpu", compute_type="int8")

---

Evaluation (Original PyTorch Checkpoint)

Evaluation metrics from the original fine-tuned model:

Metrics were computed on an NSFW-domain validation set. Performance outside the adult-content domain may degrade.

---

Intended Use

This model is suitable for:

• Japanese ASR in adult-content domains
• Offline transcription systems
• Latency-sensitive production environments
• Research on domain-specialized ASR

---

Limitations

• May produce explicit sexual content.
• No safety filtering included.
• Not appropriate for child-safe or workplace-safe applications.
• Domain shift significantly impacts WER.

---

Ethical & Safety Notice

Because the model was trained on NSFW transcripts:

• Outputs may contain explicit language.
• It may replicate biases from adult media.
• Deployments should include content moderation where appropriate.
• Access control is strongly recommended.

Users are solely responsible for compliance with applicable laws and platform policies.

---

License

Please ensure compliance with:

• The original base model license (kotoba-tech/kotoba-whisper-v2.1)
• OpenAI Whisper license terms (if applicable)
• Local laws regarding adult content processing

---

If you use this model in research, please cite the original Whisper architecture and the Kotoba Tech base model accordingly.