Optimizing Segmentation
The segmentation quality of speech-recognized subtitles directly impacts the final translation quality. This guide explains how to tune parameters for the best segmentation results.
1. Model Selection
Different languages have different optimal models:
| Language | Recommended Model | Notes |
|---|---|---|
| English | faster-whisper large-v3 | Best English recognition |
| Chinese | Qwen-ASR / FunASR | Excellent Chinese recognition |
| Japanese | openai-whisper large-v3 | Good Japanese recognition |
| Other languages | openai-whisper large-v3 | General-purpose model supporting dozens of languages |
2. Parameter Tuning
Speech Duration Parameters
| Parameter | Location | Default | Description |
|---|---|---|---|
| Max speech duration (sec) | ASR Settings | 6 | Force-split when exceeded |
| Min speech duration (ms) | ASR Settings | 3000 | Merge into adjacent subtitle if below this |
| Silence split duration (ms) | ASR Settings | 140 | Only split at silence gaps longer than this |
Tuning tips:
- Need shorter subtitles: decrease Max speech duration (e.g., 3–4 sec)
- Need longer subtitles: increase Max speech duration (e.g., 8–10 sec)
- Merge more short subtitles: increase Min speech duration (ms) (e.g., 4000)
VAD Parameters
| Parameter | Location | Description |
|---|---|---|
| VAD selection | ASR Settings | silero (default) or tenvad |
| Speech threshold | ASR Settings | Lower = more sensitive, may pick up noise as speech |
Other Parameters
| Parameter | Location | Description |
|---|---|---|
| Pre-segment audio with Whisper | ASR Settings | Recommended when using clone voices |
| Merge short subtitles | ASR Settings | Enabled by default, merges very short subtitles |
| STT again | Main UI | Re-recognize after dubbing to generate a more accurate timeline |
3. Denoise and Vocal/BGM Separation
If your video has background music or noise:
- Click More Settings...
- Enable Split Vocal/BGM
- Optionally enable Re-embed BGM to re-add the separated background audio
💡 Separating vocals significantly improves recognition accuracy, especially for videos with loud background music.
4. Single Video Interactive Mode
In single-video mode, you can pause at each stage and manually edit:
- After speech recognition → subtitle edit window appears
- After translation → translated subtitle edit window appears
- After dubbing → dubbing result edit window appears
Each window has a countdown (default 30 seconds). After timeout it auto-saves and continues. Click Stop Countdown to pause.
5. LLM Resegment
If the recognized subtitle segmentation is not ideal, use LLM Resegment:
- Enable LLM Resegment on the main interface
- Configure the LLM channel in More Settings... (supports OpenAI-ChatGPT or DeepSeek)
- Set LLM Resegment batch size (default 20 lines)
⚠️ Do not use LLM Resegment with clone voices — it disrupts the timeline.
Related Docs
- Improving AI Translation Quality — Translation mode comparison and glossary usage
- Using Local LLMs as Translation Channel — Local LLM configuration guide
- Customizing AI Translation Prompts — Custom translation prompts
- Best Settings for Video Translation — Optimal configuration for each stage
- Blank Subtitle Lines: Causes and Fixes