Skip to content

Sotto v0.3.5

Choose a tag to compare

View all tags
@github-actions github-actions released this 10 Oct 05:43
v0.3.5
645effd

Sotto 0.3.5 makes Whisper on Windows about ten times faster and lets more local models transcribe files. Dictation is pasted sooner and stops right away when you cancel. History shows what recognition, formatting and the LLM each changed.

Faster

  • Whisper on Windows is about ten times faster. Windows builds compiled whisper.cpp without compiler optimization, which slowed every Whisper model. The speeds in the model catalog were measured again on the fixed build. English-only Whisper models no longer run a language-detection pass when the language is set to auto. (#101)
  • Text is pasted sooner. A finished dictation is pasted first and saved to history afterwards, so the paste no longer waits on the database. Streaming models finish from the live preview instead of decoding the whole recording again. If the preview has a gap or the model was reloaded, Sotto falls back to the full decode. (#102)
  • Cancel stops at once. Cancelling now interrupts a running Whisper decode or a streaming model's final pass instead of waiting for it to finish. (#102)

New and improved

  • More models can transcribe files. File transcription now works with Whisper, GigaAM, Parakeet Ultra and Qwen3. Other models are refused with a message that names the ones that work. Sotto checks the model that will actually run, including after an idle unload. (#101)
  • Stages in History. A history entry shows the Recognition, formatting and LLM text in one switchable pane, along with what each step changed. The text now uses the full card width. File results use the same view, fold their details by default and can run the LLM in place. (#101)
  • Clearer LLM messages. A rejected request, for example one with a wrong model name, and an unknown provider are now described in words instead of raw codes. Skipped or rejected passes are described in words too. (#101, #102)
  • Smarter «типа» cleanup. «типа» is removed as filler in the middle of a sentence too. It is kept in «типа того», after «что-то» or «какой-нибудь», and in phrases like «поле типа string». The development dictionary now turns «Cloud Coda» into Claude Code. (#101)
  • New default color. The interface is now blue by default. Orange remains a preset, and a color you already chose is kept. Overlay templates no longer draw a rim by default. The Glow card centres its processing, pasted and error status. (#100, #101, #103)
  • Tidier windows. Pages no longer start under a blank band at the top, on both Windows and macOS. Explanations of the pipeline modes now appear as hints on hover or focus. macOS permission notices and the telemetry question share one layout. (#98, #101)

Fixed

  • Other apps' volume is restored after a crash. With Duck other audio enabled (Windows, off by default), Sotto now remembers the original volumes. If Sotto quits or crashes during a recording, those volumes are restored on the next launch. Saved levels are kept for at most seven days. (#102)
  • No more <unk> in text. Unrecognized tokens from Sherpa models are removed, including from live drafts, without leaving double spaces. (#98)
  • Microphone test errors. A paste failure no longer appears in the microphone section as a test error. (#98)
  • Less background work. History no longer reloads after every dictation while the window is hidden in the tray; it reloads once when shown. Stats refresh after the entry is saved, and the model is not loaded twice at startup. (#102)

Full changelog

Assets 18
Loading