Voice did not transcribe
How to handle voice notes where the transcript or the AI-structured fields never appear.
Voice notes go through two stages once they reach the server: the audio is transcribed to text, then the text is restructured into the description + recommended action fields on the observation. Either stage can fail. This page covers the diagnostic flow.
Symptom: transcript never appears
The audio plays back fine on the observation card but the transcript stays blank.
Three usual causes:
- The audio upload has not drained. Open the sync indicator. If the audio is still queued, the sync-stuck troubleshooter covers the recovery path. The transcription service only sees the audio after the upload completes.
- The transcription service had a transient outage. The pipeline retries with backoff; if it has run out of retries, the transcript stays blank. Reloading the workbench triggers a fresh attempt on next sync.
- The recording itself is unusable. Very short recordings (under a second), heavy background noise, or a stuck recording (where the file recorded but is silent) produce empty transcripts. Listen back; if the audio is inaudible, re-record.
Symptom: transcript appears but description / action stays blank
Stage 1 succeeded (transcript text is on the card) but stage 2 (the AI restructure) did not produce the description or recommended action fields.
This usually means the transcript is too short to restructure ("yeah this looks fine" produces neither a description nor an action), or the restructuring service had a transient blip.
Two paths:
- Type the description and recommended action manually. The transcript is still there for the audit trail; the description + action fields drive the report, and you can always type them.
- Re-record. Speak the location, then describe what you saw, then recommend an action. The structurer is much better at finding both fields when the recording sounds like an observation note rather than chatter.
Voice notes covers the speaking patterns that produce clean structured output.
Symptom: transcript looks wrong
The transcription service occasionally mishears proper nouns, unusual measurements, or technical terms. Common patterns:
- FD30S transcribes as "F D 30 S" or "FD thirty S".
- Storeys transcribes as "stories" (US English habit).
- Compartmentation is reliably correct; older systems used to mangle this.
You can edit the transcript directly on the observation card. Editing the transcript does NOT re-run the restructuring; the description and recommended action stay as they were. If you want a fresh restructure, re-record with cleaner speech.
A note on data residency
Audio bytes are uploaded to an EU-resident transcription service. The audio is retained per the agreement in your data-processing addendum (typically 7 years for housing contractors). The transcript text is stored alongside the observation in the same database region.
Data residency and GDPR covers the storage details.
When the same recording keeps failing
If a specific recording reliably fails (transcript blank, re-record produces same result), send us the audio file plus the observation id. Pre-pilot we are still tuning for the acoustics of UK building sites; pattern data helps the next release.
Where to go next
- Voice notes for the capture flow.
- Sync is stuck for the upstream upload-drain issue.
- Data residency and GDPR for where the audio is processed.