Skip to content

STT Transcripts Don't Trigger User "Speaking" State When VAD Misses Speech #3617

@philipp-eisen

Description

@philipp-eisen

Environment

livekit-agents==1.2.11

We are using the livekit session with:

...
turn_detection = MultilingualModel(),
vad=silero.VAD.load(),
...

as well as Deepgram as STT

Problem

We've noticed an issue that we at times receive a transcript from the STT model.
In traces we can see that we have a eou_detection span, however we don't have a user_speaking span and we also don't receive any user_state_changed event.
If the EOU model detects the speech as final the agent will continue in the conversation.

However, this is very problematic for us when the previous state of the user was away, because then our agent will continue to think that our the user is away and after some time end the call.

I have seen PR #3559 -- however this does not appear to change the user state.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions