Whisper cold start · make the wait observable
A healthy local STT server can still look frozen for minutes on a new user's first transcription.
A fresh OpenScreen report measured roughly 396 seconds on the first local transcription and 6.5 seconds once warm. The report identifies one-time model and Vulkan initialization and shows the same pending boolean feeding at least three loading surfaces.
This independent gate turns those timing and UI-state facts into a reproducible disclosure decision. It does not run their build, inspect their hardware or claim that a phase message fixes startup performance.
Runs in this browser · aggregate telemetry only · no recording, transcript, model file, executable path, key, private URL or customer data
Whisper cold start · timing metadata only
Turn an invisible first-run wait into one measurable product state.
Enter bounded timing and UI-state metadata only. This page accepts no recording, transcript, model file, executable path, key, private URL or customer identifier.
This gate does not run Whisper, load a model, measure accuracy, inspect hardware, prove a fix, compare audio quality or establish a paid buyer.
Readiness contract
Keep startup measurement, shared state and UX copy as separate claims.
- 01Measure one cold and one warm run on the target installation.
- 02Expose initializing versus transcribing once in the engine state.
- 03Make every loading surface consume that shared phase.
- 04Measure abandonment and first value separately from backend health.
Scope
Does this upload audio or execute Whisper?
No. It accepts only bounded timing and UI-state metadata.
Does a visible message make local transcription fast?
No. It makes a measured fixed startup cost understandable; performance work remains a separate claim.
Does the hosted CTA prove lower total latency?
No. It is an independent route that must be tested with rights-cleared audio, target geography and measured economics.