Bug Description
The Speechmatics STT plugin currently appears to hardcode transcript confidence to 1.0, so LiveKit Sessions always shows Speechmatics transcripts as 100% confident.
This makes confidence-based logic unusable, for example asking the user to repeat when ASR confidence is low.
It looks like this behavior may have been introduced in PR #4703:
Before that PR, the plugin seems to have used the reported confidence value instead of hardcoding 1.0. I do not see any explanation in the PR description/comments about why confidence was changed.
Questions:
Is hardcoding confidence=1.0 intentional?
Does the newer Speechmatics integration expose real provider confidence somewhere?
If real confidence is not available, would it be better to leave confidence unset/default instead of reporting 1.0?
Would you accept a PR that propagates Speechmatics confidence when available?
Expected Behavior
Speechmatics confidence should be propagated when available, or the plugin should avoid reporting 1.0 as if it were real confidence.
Reproduction Steps
1. Create a LiveKit agent using the Speechmatics STT plugin.
2. Log the transcription confidence from user transcript events.
3. Observe the logged confidence value or check LiveKit Sessions.
Operating System
Cloud
Models Used
No response
Package Versions
Session/Room/Call IDs
No response
Proposed Solution
Additional Context
No response
Screenshots and Recordings
No response
Bug Description
The Speechmatics STT plugin currently appears to hardcode transcript confidence to 1.0, so LiveKit Sessions always shows Speechmatics transcripts as 100% confident.
This makes confidence-based logic unusable, for example asking the user to repeat when ASR confidence is low.
It looks like this behavior may have been introduced in PR #4703:
Before that PR, the plugin seems to have used the reported confidence value instead of hardcoding 1.0. I do not see any explanation in the PR description/comments about why confidence was changed.
Questions:
Is hardcoding confidence=1.0 intentional?
Does the newer Speechmatics integration expose real provider confidence somewhere?
If real confidence is not available, would it be better to leave confidence unset/default instead of reporting 1.0?
Would you accept a PR that propagates Speechmatics confidence when available?
Expected Behavior
Speechmatics confidence should be propagated when available, or the plugin should avoid reporting 1.0 as if it were real confidence.
Reproduction Steps
Operating System
Cloud
Models Used
No response
Package Versions
Session/Room/Call IDs
No response
Proposed Solution
Additional Context
No response
Screenshots and Recordings
No response