Skip to content

Sotto 3.2.0

Pre-release
Pre-release

Choose a tag to compare

@millZach millZach released this 31 Jul 22:06

Sotto 3.2.0 simplifies the model lineup and fixes the two worst dictation-quality bugs.

Highlights

  • Two models instead of four. The lineup is now Standard (the fast English Moonshine model, bundled and ready offline) and Multi-lingual (an optional Whisper download for non-English speech). The Balanced and Accurate Whisper models are gone; installs that had them selected switch to Standard automatically. The installer is smaller because the bundled model changed.
  • Inference is always CPU/WASM. The Auto / Prefer WebGPU / CPU-WASM selector is removed — measured WebGPU inference was slower than WASM on typical hardware and the packaged runtime never shipped a WebGPU build.
  • Lists finally survive to your clipboard. The whitespace tidy pass was flattening the AI formatter's multi-line lists into one run-on line. Line breaks are now preserved, and the formatting prompt recognizes announced lists ("I want a list of things...") spoken as plain comma enumerations.
  • No more silent word loss on short segment tails. Sub-second audio segments were given a zero-token decode budget by the speech library — a guaranteed empty transcription, silently dropped. Sotto now sets an explicit budget scaled to audio length.
  • Repetition hallucinations are repaired. When the on-device model gets stuck repeating a word ("no, no, no, ..."), the loop is collapsed before the text reaches formatting, history, or paste — and the truncation guard measures against the repaired text so legitimate cleanups aren't rejected.

Install

Download and run Sotto-Setup-3.2.0.exe. Quit any running Sotto from the tray first, then install; settings, history, and downloaded models are preserved. The Standard speech model is bundled and works offline.

This build is not Authenticode-signed, so Windows may display an Unknown publisher or SmartScreen warning.

Verification

  • 1,068 unit and integration tests passed; 4 environment-gated tests were intentionally skipped.
  • TypeScript typecheck and ESLint passed.
  • Bundled Moonshine model, ONNX runtime, resources, third-party notices, and installer contents were verified, including a packaged smoke launch.

See SHA256SUMS.txt for artifact checksums.