Benchmarks

Benchmarks

Wiki · GUI · CLI · 中文

Benchmarks & Power Metrics

Measured on Apple Silicon (arm64) running a 3-minute 1080p recorded lecture clip with powermetrics hardware sampling (idle baseline ≈ 1.9 W):

Backend (--provider)	Wall Time	Avg Power (CPU / GPU / ANE)	Peak Memory	Notes
coreml + qwen3-0.6b	47 s	6.7 W / 0.2 W / 3.5 W	1.41 GB in-proc	Lowest power — Neural Engine does the heavy lifting; best on battery; zero external dependencies
coreml + whisper-turbo	87 s	15.3 W / 0.3 W / 0.4 W	1.51 GB in-proc	Whisper large-v3-turbo on CoreML; decoder mostly on CPU for short segments
gpu (llama.cpp Metal)	13 s	4.7 W / 16.0 W / —	26 MB + 3.3 GB child	Fastest; GPU bursts; needs llama-server (Qwen3-ASR 1.7B Q8)
cpu (llama.cpp)	26 s	21.2 W / 0.6 W / —	26 MB + 4.8 GB child	Universal fallback; high CPU power
api (cloud STT)	~10 s	< 1 W	negligible	Audio uploaded to provider; speed depends on network
npu (Intel Core Ultra)	16 s	NPU hardware acceleration	18 MB + 557 MB child	>6x faster than CPU on Intel Core Ultra laptops; Whisper Large-v3 Turbo

Note on the coreml default model: the measurements above were taken with the former 0.6B CoreML default. The current default qwen3-1.7b (Qwen3-ASR 1.7B MLX 8bit, running on GPU) is both more accurate and faster per upstream benchmarks (WER 1.52% vs 3.02%, RTF 0.033 vs 0.098), at the cost of roughly 2× peak memory (RSS ~2.7 GB vs ~1.4 GB) and giving up the ANE low-power path. Pick --asr-model qwen3-0.6b when battery life matters most.

👉 See the comprehensive macOS Benchmark Report for full methodology, energy breakdowns, and reproduction scripts.

---