Transcription
Subwright converts audio or video into timestamped text. Everything runs locally, with three engines: Whisper, Qwen3-ASR, and SenseVoice.
Supported formats
- Audio: MP3, WAV, M4A, FLAC, OGG, AAC, WMA
- Video: MP4, MOV, AVI, MKV, WMV, FLV, WEBM
Start transcribing
- Drag an audio/video file onto the library, or use All media → Open…
- In the Watch page, open the Script stage of the inspector
- (Optional) Adjust engine, language, model, temperature, etc.
- Click “Transcribe”
- Results stream in on the right, synced to playback
Choosing an engine
| Engine | Best for | Notes |
|---|---|---|
| Whisper | Multilingual, general material | tiny → large-v3-turbo; VAD filtering and stable timestamps |
| Qwen3-ASR | Chinese, Japanese, Korean | 0.6B / 1.7B; add the Forced Aligner for precise timing |
| SenseVoice | CJK, fast transcription | Available on macOS / Linux; not on Windows yet |
Qwen3 capability gate
- With the Forced Aligner: precise native timing; burn-in, timed export, and trim are all available.
- Without the Forced Aligner: estimated pseudo timing; transcription, burn-in, and SRT/VTT/FCPXML export still work, with a notice.
- Video trim requires precise timing; the entry point explains why when the aligner is missing.
Transcription options
| Option | Description | Recommendation |
|---|---|---|
| Engine | Whisper / Qwen3-ASR / SenseVoice | Try Qwen3-ASR for Chinese; Whisper for broad language coverage |
| Language | Auto-detect or specify source language | Specifying improves accuracy |
| Model | tiny/base/small/medium/large, Qwen3-ASR 0.6B/1.7B, SenseVoice | base for speed, large for accuracy |
| Temperature | Sampling randomness | 0 = deterministic, 1 = random, default 0 |
| Stable Timestamps | VAD-backed precise timestamps | Enable for better subtitle timing |
| Speaker Diarization | Identify different speakers | Enable for meetings |
Speaker diarization
When enabled, results are labeled Speaker 1, Speaker 2, etc. Supports up to 4 speakers.
- Requires an additional ONNX model (auto-downloaded on first use)
- Accuracy depends on audio quality and speaker count
- Different speakers are color-coded; click to jump
Word stream and resegmentation
- Word-level timestamps are stored in a local word stream during transcription
- Changing the line-length limit re-segments the current transcript without decoding again
- Works with Whisper and Qwen3; older transcripts without a stream say so honestly
- Subtitle boundaries can snap to scene changes with an adjustable range
Resume and retry
- Long transcriptions resume from a checkpoint after an interruption or crash
- Click “Abort” anytime; already transcribed content is preserved
- Adjust options and re-transcribe the same file
Shortcuts
Cmd/Ctrl + T: Start transcriptionCmd/Ctrl + .: Abort transcription
Last updated