Phase 1 · Local model quality test

Bingo Audio Test v0.6.3

Whisper runs locally. This patch repairs repeated calls when one rendition is clear and the caller's second rendition is partially garbled.

Checking device…

1. Load the local speech model

WASM remains the recommended iPhone backend. The model may be cached after the first successful load.

Compute backend Whisper model

Model not loaded.

2. Record a benchmark clip

Say several calls twice, such as “I eighteen, I eighteen; N thirty-one, N thirty-one.” Voice enhanced requests automatic gain and speech cleanup.

Clip length Microphone processing

StatusIdle

3. Transcribe and evaluate

Expected calls for this clip

Enter each intended bingo call once. The parser can now use the repeated letter as evidence when only one of the two spoken numbers is recognized cleanly.

Processing time—

Audio length—

Realtime factor—

Expected recovered—

Audio signal—

Transcript quality—

Raw Whisper transcript

No result yet.

Parsed call events

No calls parsed yet.

Elks repeat status

No repeated calls evaluated yet.

4. Parser-only test

Paste a Whisper transcript here to test the bingo parser without recording again.