Pipelines¶
WhisperJAV offers multiple processing pipelines, each trading speed for accuracy. Choose based on your content and hardware.
Pipeline Comparison¶
| Pipeline | Backend | Scene Detection | VAD | Speed | Accuracy | GPU Memory |
|---|---|---|---|---|---|---|
| Faster | Faster-Whisper | No | No | Fastest | Good | ~2 GB |
| Fast | Whisper | Yes | No | Fast | Better | ~4 GB |
| Balanced | Whisper | Yes | Yes | Medium | Best (Whisper) | ~4 GB |
| Fidelity | Whisper | Yes | Full | Slow | Maximum | ~6 GB |
| Transformers | HuggingFace | Yes | Yes | Medium | Good | ~4 GB |
| Qwen3-ASR | Qwen3 | Assembly | Assembly | Medium | Excellent text | ~4-8 GB |
| ChronosJAV | anime-whisper / Kotoba | TEN VAD | TEN VAD | Medium | Best for anime/JAV | ~4-8 GB |
Which Pipeline Should I Use?¶
| Scenario | Recommended Pipeline |
|---|---|
| First time, just want subtitles | Balanced (default) |
| Processing many files quickly | Faster |
| Anime or JAV content | ChronosJAV with anime-whisper |
| Maximum accuracy, don't mind waiting | Ensemble: Balanced + Qwen3-ASR with Smart Merge |
| No GPU / CPU only | Faster with CPU-only mode |
| Apple Silicon Mac | Transformers (MPS acceleration) |
Ensemble Mode¶
Run two passes with different pipelines and merge the results. See Ensemble Mode for details.
Specialized Pipelines¶
- ChronosJAV — anime-whisper and Kotoba models for anime/JAV content
- Qwen3-ASR — alternative ASR engine with strong Japanese text quality