Repository navigation
Conversation
… switches _ParticipantStreamTranscriptionOutput.set_participant replaced the participant identity and track, then scheduled flush() and synchronously reset the segment state. The flush task later read the new identity, track and segment id, so an in-flight final was published as an empty packet under the new participant with a fresh segment id, and the previous participant's segment was never finalized. The non-delta user transcription sink takes that path on a warm-transfer switch mid-utterance. The legacy sink had the same deferred-state race, and kept the previous participant's track because it only re-resolved the track when _track_id was empty; its final went out as the new identity on the old track. Finalize the open segment before replacing the participant, and snapshot the segment id, identity, track and latest text into the flush task so the deferred publish uses the state the segment was captured under. Always re-resolve the legacy track for the newly linked participant. Add a regression test that switches participants while an STT interim is open and asserts both sinks close the previous segment exactly once with the original sender, track and segment id.
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| # finalize the open segment for the participant it was captured for before the | ||
| # identity and track are replaced; flush() snapshots them for its publish task | ||
| self.flush() |
There was a problem hiding this comment.
🔴 Participant switches mix two users' transcripts
When set_participant runs during the first capture_text stream request, flush() cannot close its still-unassigned writer. If that capture resumes before the next chunk, the new user's text enters the old user's transcript.
Learn more
A delta transcript keeps one open text writer for all its chunks. capture_text marks the segment as capturing before awaiting writer creation, then assigns _writer after the request completes. If a participant switches during that await, flush() sees _writer=None, schedules a close that does nothing, and clears the capturing flag. The pending request can then assign the old writer; the next capture resets segment state but reuses that writer, mixing participants.
Example: A's first chunk awaits stream_text. The room switches to B, and the flush runs without a writer. A's request finishes; B's first chunk then writes to A's open stream instead of starting B's stream.
Recommended fix: Serialize participant switches with outstanding capture_text operations, or retain and close a pending writer after its creation before allowing the new participant's capture to use _writer. Cover a switch while stream_text is suspended in a delta-stream test.
Was this helpful? React with 👍 or 👎 to provide feedback.
This branch has not been deployed
1 participant
fix(voice): keep the in-flight final transcript when the participant switches mid-utterance
What breaks
RoomIO's transcription outputs finalize asynchronously:capture_text()queues aflush()task. When the local participant switches mid-utterance, the switch replaces the represented participant/track state before the queued flush runs, so the user's in-flight final transcript is published under the new participant:sender=user-b,TR_b,chunks=['']— the original text lost),Fix
Both sinks now flush before replacing the participant identity/track, and each flush snapshots the attribution it must publish with:
flush()snapshots(sender_identity, track_id, segment_id, latest_text)into_flush_task;_create_text_writeraccepts explicit sender/segment overrides, so the deferred publish never reads reset/new-participant state.flush()snapshots(represented_by/identity, track_id)into_publish_transcription;set_participantclears_track_idand always re-resolves the new participant's mic track, so the oldtrack_sidcannot leak.Tests
tests/test_room_io.py::test_participant_switch_mid_utterance_keeps_the_final_transcript(deterministic_FakeRoom/_FakeWriterharness): captures an STT interim foruser-a, switches touser-bbefore the final, thenaclose()awaits both flush tasks.user-b(assert 'user-b' == 'user-a').TR_a; no final under the new participant.test_room_io.py61 passed, transcription group 483,test_agent_session110; ruff/mypy clean on room_io.Independent review probes covered A→B, A→B→C, A→None and A→B→A: exactly one final per captured segment with correct identities/ids (no double flushing); the legacy no-mic path and
publish_on_behalfresolution are preserved. Prior art: closed #7531 fixed only the legacy sink, so the primary stream-sink bug remained. No overlap with open PRs #7933 / #7928 (different files).