Pilot results · interpretation corrected 6 September 2026
The same 20 clean synthetic fixtures, decoded three times per lane on one M1 Max MacBook Pro using Yaps 2.3.2129. Core ML Parakeet and Whisper Base are model configurations inside Yaps, not competitor products.
| Yaps output lane | Role | Runs | Words, with repeats | Micro WER | Deletion rate | Median processing | p95 processing | Maximum processing |
|---|---|---|---|---|---|---|---|---|
| Core ML Parakeet + local cleanup | Cleanup on | 60 | 1,473 | 3.46% | 0.41% | 1,124 ms | 2,807 ms | 2,958 ms |
| Core ML Parakeet, cleanup off | Cleanup off | 60 | 1,473 | 3.67% | 0.20% | 250 ms | 722 ms | 1,863 ms |
| Whisper Base, cleanup off | Cleanup off | 60 | 1,473 | 6.52% | 0.41% | 844 ms | 1,669 ms | 2,189 ms |
Processing time includes recognition, pipeline finalization and local cleanup when enabled. It excludes initial runtime construction, fixture loading, live recording, shortcut dispatch and cursor insertion. All lanes score final evaluator output; cleanup-off lanes can still apply vocabulary and formatting changes.
Latency uses 60 equally weighted decode samples per lane and R-7 linear quantiles. The 20 fixtures contain 491 reference-word positions per pass, counted three times in the 1,473-word denominator. Repeats are not independent utterances. Quality-threshold failures remain included: 3 with Core ML cleanup on, 3 with Core ML cleanup off and 6 with Whisper.
Interpretation corrected 6 September 2026. Original numbers are unchanged. Read the evidence audit, external references, remaining limitations and reproduction instructions.
“Micro” means edit-operation counts and reference words were summed before calculating each rate. It is not the macro average of per-clip percentages.
Whisper Base with cleanup off had 2.85 percentage points higher WER, 0.20 points higher deletion rate, a 595 ms higher median, and a 947 ms higher p95 than Core ML Parakeet with cleanup off in these matched pilot runs.
The Core ML cleanup-enabled lane includes local semantic cleanup. Its 1,124 ms median and 2,807 ms p95 describe recognition and finalization together, before insertion into another application.
The sanitizer retained counts and latency while excluding transcripts, references, prompts, and application context. Public file integrity is recorded with SHA-256 digests.
Human speech, multiple accents and languages, live microphone capture, controlled thermal order, network observation, more devices, and third-party systems are still required.
The corpus, device matrix, order, and thermal procedure were not preregistered. These observations describe these sequential runs, on this corpus, device, software version, and date.
Download the sanitized rows, aggregates, manifests, and checksums.