Pipeline
What happens to a recording, step by step
Processing runs on the server. You can leave the page and come back; every stage reports its own state.
Simulation with sample data. Scroll: the line travels with you
- Preparing the recordingThe line is here
The file is checked before anything else
Size, file signature and real audio duration are verified. The original recording is kept as it is; processing works on a checked copy.
- Speech to textThe line is here
Whisper turns speech into a transcript
The language is the one you set for the recording, or it is detected. The transcript keeps timestamps so later stages can point back to the conversation.
- Separating speakersThe line is here
Planned: turns are assigned to speakers
Speaker separation is planned. Today the stage is marked unavailable with the reason, and the rest of the pipeline continues on the text.
- Summary and analysisThe line is here
A summary with evidence
The analysis stage writes a summary, action items and insights, each linked to the transcript lines behind it. QC scoring, problem detection and coaching are planned. Test-provider output is labeled and never presented as a real analysis.
- ReviewThe line is here
Your team works in the console
Overview, call results, transcript, search, share links and exports. Personal-data masking and access rules apply to every view.
- Truthful statesThe line is here
A missing stage is shown, not filled in
Each stage of a job carries an explicit state. If a stage stops, the report shows why, in the language of the console.
- 1Preparing the recordingDone
- 2Speech to textDone
- 3Separating speakersUnavailable
- 4Summary and analysisRunning
- Not started:
- The stage has not been reached yet.
- Waiting:
- The recording is in the processing queue.
- Running:
- The stage is being processed right now.
- Partial:
- Processing finished with gaps; the stages show what is missing.
- Unavailable:
- The stage could not run, for example when speaker separation is not available for this recording.
- Failed:
- The last attempt failed with a named reason. Automatic retries are bounded and visible.
- Where the AI runsThe line is here
Every stage checks the workspace policy first
Each workspace has an AI policy: which providers, regions and models may be used for speech to text, speaker separation and analysis. The policy is checked before a stage runs. A provider the policy does not allow is never called and never used as a fallback.
- Local WhisperAllowed
- Local analysis modelAllowed
- Cloud provider outside the policyRefused