Description
When recording a video call through speakers, Quill transcribes much of the remote participant's speech from both system.caf and mic.caf. The resulting transcript contains adjacent, nearly identical segments labeled first as them and then as me.
This makes the two-track speaker labeling unreliable and approximately doubles large portions of the transcript.
I understand that mic_voice_processing can enable Apple's echo cancellation. However, with the default configuration, the transcript can silently contain extensive duplicate speech without warning or post-processing.
Environment
- macOS: 26.6
- Mac: 2021 MBP
- Quill commit/version:
7ea94e27cbf403dd1fa14f460b1bea56bf78477d
- Transcription engine:
parakeet-tdt-0.6b-v2-coreml
- Audio output: Mac speakers
mic_voice_processing: false / not configured- Call application: Google Meet (browser)
Example
"them" said:
Thanks so much for joining and being open to a conversation.
The transcript emitted it twice:
**[0:44] them:** Thanks so much for joining and being open to a conversation.
**[0:44] me:** Thanks so much for joining and being open to a conversation.
The same pattern continues through much of the recording.
In a 23-minute recording, there were approximately:
- 371 total transcript entries
- at least 95 adjacent near-identical cross-speaker pairs
- roughly 51% of entries involved in those duplicate pairs
There are also cases where the duplicate mic transcription differs slightly, producing conflicting text or absorbing part of a neighboring utterance.
Description
When recording a video call through speakers, Quill transcribes much of the remote participant's speech from both
system.cafandmic.caf. The resulting transcript contains adjacent, nearly identical segments labeled first asthemand then asme.This makes the two-track speaker labeling unreliable and approximately doubles large portions of the transcript.
I understand that
mic_voice_processingcan enable Apple's echo cancellation. However, with the default configuration, the transcript can silently contain extensive duplicate speech without warning or post-processing.Environment
7ea94e27cbf403dd1fa14f460b1bea56bf78477dparakeet-tdt-0.6b-v2-coremlmic_voice_processing:false/ not configuredExample
"them" said:
The transcript emitted it twice:
The same pattern continues through much of the recording.
In a 23-minute recording, there were approximately:
There are also cases where the duplicate mic transcription differs slightly, producing conflicting text or absorbing part of a neighboring utterance.