diff options
| author |  sommerfeld <sommerfeld@sommerfeld.dev> | 2026-05-13 13:43:23 +0100 |
|---|---|---|
| committer | sommerfeld <sommerfeld@sommerfeld.dev> | 2026-05-13 13:43:23 +0100 |
| commit | d643aaf8a3a3fc8a31f39cb2ae3eaaf681300d4e (patch) | |
| tree | d3a030b6955093707d095fbe37a9c5ef1137ae93 /meta | |
| parent | 5c241d65ed4a6ec2bc3e5d75d6858ed6722f1b17 (diff) | |
| download | dotfiles-d643aaf8a3a3fc8a31f39cb2ae3eaaf681300d4e.tar.gz dotfiles-d643aaf8a3a3fc8a31f39cb2ae3eaaf681300d4e.tar.bz2 dotfiles-d643aaf8a3a3fc8a31f39cb2ae3eaaf681300d4e.zip | |
perf(dictate): switch default model to base for ~5x speedup
large-v3-turbo-q5_0 ran ~1-2x realtime on the T490's CPU, making
push-to-talk feel sluggish. The base multilingual model is ~142 MB
(vs 547 MB) and runs ~7-10x realtime, dropping perceived latency on
short utterances from a few seconds to near-instant.
Quality on short EN/PT dictation remains usable; bump WHISPER_MODEL
to small or large-v3-turbo if accuracy matters more than latency.
Diffstat (limited to 'meta')
| -rw-r--r-- | meta/extra.txt | 4 |
1 files changed, 3 insertions, 1 deletions
diff --git a/meta/extra.txt b/meta/extra.txt index fb87c3d..36c311f 100644 --- a/meta/extra.txt +++ b/meta/extra.txt @@ -10,5 +10,7 @@ tesseract-data-eng tesseract-data-por # Speech-to-text (used by ~/.local/bin/dictate) +# `base` multilingual: ~142 MB, ~7-10x realtime on a 4c CPU. Override +# WHISPER_MODEL in the script's environment to use a different ggml model. whisper.cpp -whisper.cpp-model-large-v3-turbo-q5_0 +whisper.cpp-model-base |