Apple SpeechAnalyzer vs whisper.cpp on Mac: Reproducible ASR Benchmark

Four complete speech-recognition benchmark runs over the same deterministic 40-speaker LibriSpeech test-clean snapshot. The deposit includes 160 clip-run rows, complete raw results, the corpus manifest, citation metadata, checksums, and a dependency-free verifier. Apple SpeechAnalyzer produced 1.98% word error rate and whisper.cpp small.en produced 4.28% on this snapshot. Repeated post-speech latency overlapped: Apple median 125–132 ms and p95 194–201 ms; whisper.cpp median 122–125 ms and p95 152–161 ms. Important limitation: this is clean read English audiobook speech on one Apple M5 Max running macOS 26.5. It is not ordinary desktop dictation, noisy-speech validation, broad accent coverage, multilingual evaluation, accessibility validation, or a population study. Canonical study and methodology: https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-macPublic source artifacts: https://github.com/mvplab-ai/mac-asr-benchmark