Product docs
Transcription Parameters by Engine
Map tunable settings across Whisper file, Whisper realtime, realtime models, and advanced ASR—with scene presets and plan gates.
Transcription Parameters by Engine
Advanced
Advanced transcription tuning screenshot
What this page is for
Advanced settings differ for Whisper file, Whisper realtime, realtime models, and advanced ASR (ML Engine tab) routes. Mixing them up causes “docs mention a knob I cannot see.”
This is the hub page—read Advanced Parameter Transcription for tuning workflow first.
Plans (summary)
| Capability | Free | Standard | Pro |
|---|---|---|---|
| Advanced file parameters | No | Yes | Yes |
| Realtime presets | — | Yes | Yes |
| Realtime custom params | — | No | Yes |
| Advanced ASR models | — | — | Yes |
Whisper file transcription
When: offline batch workflows (files, links, watch folders) with a Whisper model.
Typical groups (visible with Standard+ advanced settings):
| Group | Why tune |
|---|---|
| Scene presets | One-click VAD/decoding bias (table below) |
| VAD / no-speech thresholds | Fewer hallucinations, cleaner cuts |
| Decoding | Greedy vs beam, best-of, context |
| Segments & offsets | Partial files, length limits |
| Parallel & separation | Pro features such as diarization (separate toggles) |
Whisper file scene presets
| Preset | Best for | What moves |
|---|---|---|
| General | Most media | Balanced VAD/decoding |
| Dialogue | Interviews, two-speaker | Shorter segments |
| Speech | Monologue | Longer segments, more context |
| Meeting | Multi-speaker | Meeting-oriented VAD |
| Course | Lectures | Long segments, large context |
| Noisy | High background noise | Stronger VAD/denoise bias, beam |
| Media | Music + dialogue | Conservative no-speech, light denoise |
| Custom | Team template | Saved manual baseline |
Whisper realtime
When: mic/app/global realtime with Whisper as the engine.
Not the same table as file mode. Common areas:
- Step/window sizing (latency vs stability)
- Segmentation mode (when lines appear)
- Silero / energy VAD thresholds
- Thread count
Stabilize RTF before chasing perfect wording. GPU can help but is not mandatory everywhere.
Realtime models
When: realtime workflows with models from the Realtime models group (e.g. Zipformer).
Focus on VAD and chunking, not Whisper beam search:
- VAD scene presets
- Min speech / silence durations
- Min / max segment length
- Padding, merge gaps
- Threads
Do not apply Whisper file “no-speech / beam” advice here.
Advanced ASR (ML Engine tab)
When: Pro file jobs with a model from ML Engine.
- Languages and limits vary by model—see Advanced ASR models.
- Advanced fields live on a separate form—do not merge with Whisper VAD tables.
- Offline file focus today; realtime-model knobs do not apply.
If the UI disagrees with this page
Ship cadence can add or hide controls. Trust the app UI and send feedback. Engineering uses whisper-form.factory visibility and CLI mappers as source of truth (marketing-content-audit.md in the repo).
FAQ
Can Standard customize realtime params?
Presets yes; custom realtime configuration requires Pro.
Do file presets affect link jobs?
Whisper file parameters apply to supported offline batch entry points.
Where is GPU?
GPU is an engine/runtime choice—see GPU Transcription.
