English meaning
Latest paired comparison: 676 saved WAVs and gold transcripts. RambleFix's shipped Apple Neural Engine path and Handy are statistically tied on meaning.
Benchmark method
Same audio, local engines, honest caveats.
RambleFix benchmarks replay saved WAV recordings through local engines — every engine hears the exact same audio — and compare outputs against Gemini-cross-checked gold transcripts. Cloud models are used only to confirm gold labels, not in the product path. “Meaning” is the share of what was said that survives into the pasted text; we use it instead of word error rate because RambleFix strips fillers on purpose, and word error rate would count that as a mistake. Engines tested alongside RambleFix’s router include Whisper large-v3-turbo (Apple MLX build), Qwen3 ASR (MLX), and Nvidia Parakeet. Full per-engine tables are in docs/bakeoff_primary_engine_20260712.md.
Engine speed
Across 226 paired recordings, warm median decode was 153ms for RambleFix vs 411ms for Handy: 2.68× measured and reported conservatively as 2.6×. Model loading was excluded for both. On a separate set of 40 saved English clips, RambleFix decoded 7.2× faster than whisper.cpp small.en (153ms vs 1,105ms).
Voice input context
Wispr Flow cites 45 wpm keyboard vs 220 wpm Flow. Willow cites 150 wpm speech vs 40 wpm typing, and roughly 35 minutes saved per day for 2,000 words. We use those as public market context, not as a RambleFix-specific productivity guarantee.
Hindi+English
13 mixed-language clips. RambleFix meaning 0.885 vs 0.661-0.702 for tested local general engines.
Claim boundary
The 2.6× and 7.2× figures are warm engine decode, not release-to-paste app latency. whisper.cpp is an engine proxy for Whisper-based apps, not an app-level OpenWhispr comparison. Paid cloud comparisons remain directional. The Hindi+English comparison uses 13 clips.
This page is the public summary. The internal regression harness also checks blank/no-speech guards, structure safety, protected terms, local-only runtime rules, and release-to-paste timing before a public build. Productivity context comes from public Wispr Flow and Willow claims; RambleFix-specific measured claims are the local engine speed, meaning, and offline-runtime results above.
Sources: Wispr Flow, Willow writer tools, and Willow document tools.