IRAVOICE RESEARCH BRIEF
Apple SpeechAnalyzer vs whisper.cpp on Mac
Published July 26, 2026 · Version 1.0.0 · CC BY 4.0

CANONICAL RESEARCH PAGE
https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-mac

DIRECT RESULT
On a deterministic 40-speaker LibriSpeech test-clean snapshot, Apple
SpeechAnalyzer produced 1.98% word error rate and 1.02% character error rate.
whisper.cpp 1.8.4 with ggml-small.en produced 4.28% word error rate and 1.79%
character error rate.

Across two complete final-commit runs per engine:
- Apple median post-speech latency: 125–132 ms; p95: 194–201 ms.
- whisper.cpp median post-speech latency: 122–125 ms; p95: 152–161 ms.
- Every run completed 40/40 clips with zero failures.
- Accuracy matched exactly across each engine's repeated runs.

WHAT WAS TESTED
- Apple SpeechAnalyzer progressiveTranscription through IraVoice's production
  streaming adapter.
- whisper.cpp server 1.8.4 with ggml-small.en through IraVoice's production
  batch adapter.
- 40 clips: one deterministic clip from each LibriSpeech test-clean speaker,
  balanced 20 female and 20 male speakers.
- 607 reference words per run and 247.025 seconds of audio per run.
- 100 ms chunks fed at real-time pace after one unreported warmup clip.
- Apple M5 Max, Mac17,7, 128 GiB RAM, macOS 26.5 build 25F71.

WHAT WAS NOT TESTED
This is clean read English audiobook speech, not spontaneous desktop dictation,
noisy rooms, broad accent coverage, accessibility validation, multilingual
recognition, or a population study. It covers one Mac, OS build, locale, and
Whisper model. It does not compare Superwhisper, Wispr Flow, Raycast, or other
commercial dictation products.

REPRODUCIBILITY
Public artifacts, manifest, checksums, and scoring verifier:
https://github.com/mvplab-ai/mac-asr-benchmark

Tested proprietary application revision:
81ecbb9582d0990827c76be0e685124d2be651c8

Versioned GitHub release:
https://github.com/mvplab-ai/mac-asr-benchmark/releases/tag/v1.0.0

Manifest SHA-256:
171639597cda2a98311fab77b169aa7086e95321f402b50c1f93b597dbd7fc36

The production IraVoice runner remains proprietary. The public repository
contains the complete benchmark artifacts and an executable verifier that
recomputes the published aggregate metrics from clip-level edit counts.

Full JSON:
https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-mac.json

Clip-level CSV:
https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-mac.csv

Public Hugging Face mirror:
https://huggingface.co/datasets/researchaudio/apple-speechanalyzer-vs-whisper-cpp-mac

PREFERRED CITATION
IraVoice for Mac. Apple SpeechAnalyzer vs whisper.cpp on Mac: Reproducible ASR
Benchmark. Version 1.0.0, July 26, 2026.
https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-mac

CONTACT
Use the contact route published at https://iravoice.com/press