Datasets:
File size: 4,981 Bytes
57b9770 527b696 57b9770 527b696 57b9770 b63a55d 527b696 b63a55d 57b9770 527b696 57b9770 527b696 57b9770 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 | ---
license: cc-by-4.0
language:
- en
task_categories:
- automatic-speech-recognition
tags:
- benchmark
- macos
- apple-speechanalyzer
- whisper-cpp
- word-error-rate
size_categories:
- n<1K
pretty_name: Apple SpeechAnalyzer vs whisper.cpp on Mac
configs:
- config_name: default
data_files:
- split: test
path: data/clip-results.csv
---
# Apple SpeechAnalyzer vs whisper.cpp on Mac
Four complete speech-recognition benchmark runs over the same deterministic
40-speaker LibriSpeech test-clean snapshot:
| Engine | Model path | WER | CER | Repeated median post-speech latency | Repeated p95 |
| --- | --- | ---: | ---: | ---: | ---: |
| Apple SpeechAnalyzer | `progressiveTranscription` on macOS 26.5 | 1.98% | 1.02% | 125–132 ms | 194–201 ms |
| whisper.cpp server | 1.8.4 · `ggml-small.en` | 4.28% | 1.79% | 122–125 ms | 152–161 ms |
Every run completed 40/40 clips with no failures. Accuracy matched exactly
across the two repeated runs for each engine.
Canonical research page, readable methodology, visual results, and correction
path:
**https://iravoice.com/research/apple-speechanalyzer-vs-whisper-cpp-mac**
Public artifact repository, verifier, and disclosed production revision:
**https://github.com/mvplab-ai/mac-asr-benchmark**
Versioned GitHub release with downloadable artifacts:
**https://github.com/mvplab-ai/mac-asr-benchmark/releases/tag/v1.0.0**
## What is included
- `data/clip-results.csv`: 160 clip-run rows with engine and model labels,
source commit, manifest digest, hardware, OS, clip and speaker identifiers,
timing, WER, CER, word edit counts, references, and hypotheses.
- `raw/full-results.json`: all four native run records, including complete
engine, environment, protocol, summary, and clip-level output.
- `manifest.json`: the deterministic 40-clip selection and reference text.
- `CITATION.cff` and `benchmark.bib`: citation metadata.
- `CHECKSUMS.sha256`: hashes for every mirrored publication artifact.
- `source-provenance.json`: the tested private-application revision and
visibility disclosure.
- `verify.mjs`: a dependency-free verifier that recomputes aggregate WER and
CER from the clip-level edit counts and checks repeated-run invariants.
The original LibriSpeech audio is not mirrored here. Obtain the official
test-clean archive from [OpenSLR SLR12](https://www.openslr.org/12). The public
GitHub repository includes the deterministic manifest, raw outputs, checksums,
and an executable scoring-consistency verifier. The tested application revision
is preserved in every run record. IraVoice's production application source
remains proprietary, so the public repository does not claim that the engine
runner itself is open source.
## Protocol
- Corpus: one deterministic clip from each of the 40 LibriSpeech test-clean
speakers; 20 female and 20 male speakers; 8–20 reference words and 4–10
seconds per clip.
- Hardware: Apple M5 Max, `Mac17,7`, 128 GiB memory.
- OS: macOS 26.5 build 25F71.
- Audio: mono 16 kHz PCM fed in 100 ms chunks at real-time pace.
- Warmup: the first clip is run once and excluded, then all 40 clips are
measured.
- Engines: the production IraVoice Apple SpeechAnalyzer streaming adapter and
whisper.cpp batch-server adapter.
- Accuracy: micro-averaged WER and CER after disclosed normalization. Digits are
not rewritten as words. No vocabulary injection, formatting model, cleanup,
or hand correction is applied.
- Latency: measured after the final audio chunk until the complete raw
transcript returns. Startup, formatting, and text insertion are excluded.
- Repetition: two complete final-commit runs per engine.
## Limitations
This is clean read English audiobook speech, not spontaneous desktop dictation.
The 40 clips are a deterministic speaker-balanced snapshot, not all 2,620
test-clean clips. The results cover one Mac, OS build, locale, and Whisper
model. They do not establish noisy-speech, accented-speech, multilingual,
accessibility, disability, or population-wide performance.
Apple SpeechAnalyzer is tested as a streaming engine and this whisper.cpp path
as a batch server. Only the disclosed final transcript and post-speech interval
are compared. Repeated latency ranges were close; this dataset does not support
a decisive universal latency winner.
IraVoice publishes this study and currently ships the Apple recognition path.
Inspect the code, manifest, raw outputs, and limitations before reusing the
headline result. The study did not test Superwhisper, Wispr Flow, Raycast, or
any other commercial dictation product.
## 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
The published result data and metadata are licensed CC BY 4.0. LibriSpeech
source material remains subject to the corpus license and attribution described
by OpenSLR.
|