fix: prevent VAD from driving user_state when turn_detection=sst

[MdSadiqMd]

Closes #5580

Summary

Added user_state_source configuration to TurnHandlingOptions with three modes: "vad", "stt", and "auto" (default). Implemented _vad_drives_user_state property in AudioRecognition that encapsulates the decision logic. VAD can now run for interruption detection without affecting user_state when user_state_source="stt", solving false positives from background noise. Fully backward compatible with default "auto" mode

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 3 additional findings.

[miguelmoralai]

I think you should implement Option B described in the issue for a more reliable solution:

Option B — explicit configuration. Add a user_state_source: Literal["vad", "stt", "auto"] field to TurnHandlingOptions (or AgentSession). "auto" keeps current behavior. "stt" makes the VAD branch skip the _speaking writes while still running VAD inference for interruption detection. "vad" is today's default

[MdSadiqMd]

I think you should implement Option B described in the issue for a more reliable solution:

Option B — explicit configuration. Add a user_state_source: Literal["vad", "stt", "auto"] field to TurnHandlingOptions (or AgentSession). "auto" keeps current behavior. "stt" makes the VAD branch skip the _speaking writes while still running VAD inference for interruption detection. "vad" is today's default

I thought Option A might be a good fit, as it just acts as a fix for the existing system, After thinking Option B makes more sense for long term, thus making the change now

[miguelmoralai]

I think you should implement Option B described in the issue for a more reliable solution:

Option B — explicit configuration. Add a user_state_source: Literal["vad", "stt", "auto"] field to TurnHandlingOptions (or AgentSession). "auto" keeps current behavior. "stt" makes the VAD branch skip the _speaking writes while still running VAD inference for interruption detection. "vad" is today's default

I thought Option A might be a good fit, as it just acts as a fix for the existing system, After thinking Option B makes more sense for long term, thus making the change now

Yep but Option A (the one you implemented) has trade offs. I mean selecting vad/stt as turn_detection does not mean you also want to handle user_state through the same method. IMO seems a more logical but still wrong assumption. Ideally the user should be able to select both

[MdSadiqMd]

@miguelmoralai, can you please verify the changes

[miguelmoralai]

@claude review

[MdSadiqMd]

Looks like claude is not up

cc: @miguelmoralai

[MdSadiqMd]

Bump @miguelmoralai