]> git.giacomolli.uk - onechat/commit
Benchmark local voice transcription
authorAnderson A. Giacomolli <redacted>
Wed, 11 Nov 2026 02:07:16 +0000 (23:07 -0300)
committerAnderson A. Giacomolli <redacted>
Wed, 11 Nov 2026 02:07:16 +0000 (23:07 -0300)
commit23d56510bff3cda373977f2e85360b2d4ef8e7e2
treeda2ff3b2fdc1c1b7d4ea6a295d7b920d21640ecf
parent12bd9e96f3b3e67290484e3829461e054110db27
Benchmark local voice transcription

A separate Nix package compares two pinned Whisper models on public
English and Portuguese Ogg Opus excerpts. It prints each transcript,
wall time, and peak memory so the local voice model can be selected
without an external service or private audio. A decode probe selects 16
kHz mono WAV conversion when direct Opus input is not supported.
flake.nix
nix/voice-benchmark.nix [new file with mode: 0644]
nix/voice-model.nix [new file with mode: 0644]