Benchmark results

Recorded on the primary development machine (Ryzen 5900X-class, RTX 3070 Ti, Hyprland Wayland).

Fixtures

Fixture	Duration	Notes
Long speech sample	~50 s	CUDA transcribe-only numbers below
Setup fixture	~9.9 s	CPU and CUDA dictation-path benches

bench dictation reads an existing WAV and does not delete the source file.

CUDA profile (power-user-nvidia)

CUDA skaldd build, warm ggml-large-v3-turbo-q5_0.bin.

Transcribe only (skald bench end-to-end)

Long fixture (~50 s):

Run	Audio	Model load	Transcribe	Total ASR
Warm	49962 ms	0 ms	614 ms	614 ms
Warm (repeat)	49962 ms	0 ms	607 ms	607 ms

Short fixture (~4.9 s):

Run	Audio	Model load	Transcribe	Total ASR
Warm	4906 ms	0 ms	263 ms	263 ms

Cold model load (skald bench model-load after unload): 125–139 ms.

Full dictation path (skald bench dictation)

Short fixture (~4.9 s), cleanup disabled (--no-cleanup):

Run	Transcribe	Stop-to-clipboard	Cleanup
1	233 ms	287 ms	no
2 (repeat)	235 ms	287 ms	no

With cleanup (--cleanup, OpenRouter):

Transcribe	Cleanup	Stop-to-clipboard
250 ms	815 ms	1118 ms

Paste attempt (--paste --no-cleanup) from an unfocused terminal: paste skipped (active target unstable); stop-to-clipboard 322 ms. Re-run with a stable editor focused to measure stop-to-insert.

CUDA sign-off

Field	Value
Machine	Ryzen 5900X-class + RTX 3070 Ti
Session	Hyprland Wayland
Profile	power-user-nvidia
ASR model	ggml-large-v3-turbo-q5_0.bin
Validated	2026-06-11

CPU profile (cpu-safe)

CPU skaldd build, cpu-safe profile, ggml-small.en.bin, setup fixture (~9.9 s).

Transcribe only (skald bench end-to-end)

Run	Audio	Model load	Transcribe	Total ASR
Warm	9940 ms	221 ms	2718 ms	2939 ms
Warm (repeat)	9940 ms	217 ms	2679 ms	2896 ms

Cold model load (skald bench model-load after unload): 208 ms.

Full dictation path (skald bench dictation)

Setup fixture (~9.9 s), cleanup disabled (--no-cleanup):

Run	Model load	Transcribe	Stop-to-clipboard
1	0 ms	2775 ms	2891 ms
2 (repeat)	215 ms	2702 ms	3028 ms

CPU sign-off

Field	Value
Machine	Ryzen 5900X-class
Session	Hyprland Wayland
Profile	cpu-safe
ASR model	ggml-small.en.bin
Validated	2026-06-24

Latency targets

Targets assume a 10-second utterance. On the CUDA profile, warm local-only dictation (no cleanup) is well under the 1.5 s stop-to-clipboard p50 target. Cleanup adds provider latency; the short CUDA fixture run stayed under 1.2 s total stop-to-clipboard.

On the CPU profile with small.en, warm stop-to-clipboard is about 2.9–3.0 s for the ~10 s setup fixture.

Re-run locally

just bench-e2e /path/to/sample.wav
just bench-dictation /path/to/sample.wav --no-cleanup
just bench-dictation /path/to/sample.wav --cleanup
just bench-dictation /path/to/sample.wav --paste
just bench-model-load

For CPU-safe numbers, use a CPU skaldd build and skald config profile cpu-safe before benchmarking.