On-device transcription, missing audio playback, persistent synchronization issues and misleading upload status
Test environment
Device: Apple iPad Pro (2021)
Test date: November 17, 2026
Audio source: Built-in microphone
Transcription provider: On Device
Language: English
Steps to reproduce
Select On Device under Recording & Transcription.
Enable Send raw audio to Omi.
Record approximately 40 seconds of speech.
Finish recording and verify transcript, summary and audio playback.
Disable Send raw audio to Omi.
Repeat the recording test.
Observed behavior
With raw audio forwarding enabled:
Successful transcription and summarization.
Processing time of approximately 30–45 seconds for a 38-second recording.
Original audio playback available.
With raw audio forwarding disabled:
Successful transcription and summarization.
Processing time of approximately 15 seconds or less for a 33-second recording.
Live transcription during recording.
Original audio playback unavailable.
Recording interface continues to display “Uploading audio for transcription.”
Diagnostic findings
Debug logs confirm the following configuration values during the respective tests:
service_id: onDeviceWhisper
stt_service: soniox
custom_stt: enabled
forward_raw_audio_to_secondary: true (first test)
forward_raw_audio_to_secondary: false (second test)
Both configurations maintain an active WebSocket connection to api.omi.me.
Additionally, the application continues to report five missing/pending WAL segments associated with 39.26 seconds of audio, with zero segments synchronized. The oldest pending segment reached 2,457 seconds (40 minutes, 57 seconds).
Expected behavior
Users should be able to retain, play and export their original recordings locally without requiring raw audio transmission to Omi’s cloud servers.
The recording interface should also accurately distinguish local audio processing from cloud uploads.
Additional technical questions
Is the loss of audio playback intentional or a bug?
Is original audio retained locally when raw audio forwarding is disabled?
Does the Audio Bytes webhook operate independently of raw audio forwarding?
Is Soniox actually receiving audio when on-device transcription is selected?
Could the persistent WAL synchronization issue contribute to unsuccessful conversation retrieval through the ChatGPT MCP integration?
Full diagnostic logs and screenshots are available upon request.
I am extremely encouraged by the transcription performance and flexibility of Omi, but reliable audio preservation, privacy controls and external integrations are essential to its usefulness as a long-term personal knowledge management system.
Thank you for your time and for making Omi open source.
Log in to comment and vote
No comments yet
Be the first to share your thoughts.