-
Notifications
You must be signed in to change notification settings - Fork 1.4k
Open
Labels
captureLayer: Audio recording, device pairing, BLELayer: Audio recording, device pairing, BLEp1Priority: Critical (score 22-29)Priority: Critical (score 22-29)
Description
Problem
Per-hour analysis of Deepgram transcription usage (Jan 2026) shows significant cost waste from silent or near-silent streams being billed:
| Bucket | % of Total Hours |
|---|---|
| < 1 WPM (near silent) | 9.7% |
| 1-5 WPM (mostly silent) | 9.5% |
| 5-20 WPM (light use) | 15.9% |
| 20-50 WPM (normal) | 16.6% |
| > 50 WPM (active) | 48.3% |
19.2% of all transcription hours are below 5 WPM — essentially silence being billed. 9.7% are below 1 WPM (near-total silence).
- 2,456 unique users had near-silent hours (<1 WPM)
- 2,992 users had mostly-silent hours (1-5 WPM)
Proposed Fix
Auto-pause Deepgram streaming when WPM drops below threshold:
- If WPM < 1 for 5+ consecutive minutes → pause the stream
- Resume when audio activity is detected (via local VAD or similar)
Expected Impact
- Recover ~9.7-19.2% of transcription cost depending on threshold chosen
- No impact on user experience — silent streams produce no useful transcription anyway
Data Source
Deepgram usage analytics, Jan 2026, per-hour granularity. Analysis by @mon.
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
captureLayer: Audio recording, device pairing, BLELayer: Audio recording, device pairing, BLEp1Priority: Critical (score 22-29)Priority: Critical (score 22-29)