Anti-hallucination controls
Tune timestamp, context, confidence, and silence thresholds when a specific setup needs stricter decoding.
v0.11.0 Advanced Tab
Advanced Whisper.cpp Settings for Linux Dictation
Vocalinux keeps the default setup simple, but power users can unlock advanced whisper.cpp decoding controls for anti-hallucination tuning, initial prompts, and Remote API server configuration.
Remote Server options
Configure the Remote API URL, optional bearer token, and endpoint format from the same advanced area.
Resettable power-user mode
Advanced options are opt-in and can be reset when experimentation produces worse dictation.
Whisper.cpp Decoding Controls
No Timestamps
Disables timestamp token generation to reduce timestamp-related hallucinations.
No Context
Stops whisper.cpp from conditioning on previous text when past context causes error loops.
Temperature
Controls decoding randomness. The default 0.0 keeps dictation deterministic.
Temperature Increment
Sets fallback steps for retry behavior, or disables fallback entirely with -1.0.
Entropy Threshold
Helps detect repetition loops during decoding.
Logprob Threshold
Triggers fallback when average token confidence is too low.
No-Speech Threshold
Controls how strongly whisper.cpp treats audio as silence.
Initial Prompt
Adds names, jargon, punctuation style, or other context to steer transcription.
When to Use Advanced Settings
- Suppress repeated phrases or timestamp artifacts in long dictation sessions.
- Bias transcription toward project names, technical jargon, or personal vocabulary.
- Keep whisper.cpp deterministic for notes, code comments, and messages.
- Move Remote Server settings out of the way until power-user mode is enabled.
Keep Defaults Unless You Need Them
The default whisper.cpp settings are chosen for reliable everyday dictation. Advanced values can improve niche workflows, but they can also reduce accuracy if pushed too far. Change one value at a time and reset if quality drops.