30 lines
1.6 KiB
Markdown
30 lines
1.6 KiB
Markdown
# Corrections
|
|
|
|
Linnet includes an inline correction widget for fixing transcription errors. Corrections are saved locally and used to fine-tune your local Whisper model over time.
|
|
|
|
## Making a correction
|
|
|
|
1. Open a session
|
|
2. Click any transcript line to enter edit mode
|
|
3. Fix the text inline
|
|
4. Press **Enter** or click **Save**
|
|
|
|
The corrected text replaces the original in the session view. The original is preserved in the correction log for model retraining.
|
|
|
|
## How corrections improve accuracy
|
|
|
|
Each correction you make is stored in a local dataset. Linnet periodically runs a fine-tuning pass on your local Whisper checkpoint using these corrections, biasing the model toward your voice, speech patterns, and vocabulary.
|
|
|
|
This is entirely local. Corrections are never sent to CircuitForge servers.
|
|
|
|
The fine-tuning pass runs in the background when Linnet is idle. You can also trigger it manually from **Settings** under **Model** then **Retrain now**.
|
|
|
|
## What corrections do not fix
|
|
|
|
Corrections fix recognition errors (wrong words, missed words). They do not adjust tone labels — tone annotation is based on audio characteristics, not the text content. If a tone label is wrong, that reflects model uncertainty about the audio segment, not a transcription error.
|
|
|
|
Tone label feedback is on the roadmap as a separate feature.
|
|
|
|
## Resetting corrections
|
|
|
|
If you want to discard your local correction dataset and start fresh from the base model, go to **Settings** then **Model** then **Reset correction data**. This removes all saved corrections and resets to the bundled checkpoint.
|