Windows 1.9.0
·
276 commits
to main
since this release
Highlights
- August 2026 model refresh: OpenAI gpt-transcribe and gpt-live-transcribe, AssemblyAI Universal-3.5 Pro, Claude Sonnet 5, Gemini 3.6 Flash, Gemini 3.5 Flash Lite, Grok 4.5, Mistral Medium 3.5 and Qwen 3.6 27B.
- Deepgram live dictation fixed: streaming always timed out after 10 seconds, and
SpeechStarted/UtteranceEnd/ endpointing frames failed to parse, dropping words mid-sentence.
Features & Improvements
- Remove filler words in live dictation: the setting previously applied only to non-streaming transcription.
- More reliable streaming shutdown: wider WebSocket close-code handling, a safety net so a stopped recording always reaches a terminal status, and a fixed startup race when recording immediately after launch.
Bug Fixes
- Long Groq dictations no longer fall back to the raw transcript: Groq applies an undocumented 2,048 completion-token default, so long transcripts came back
finish_reason=lengthand were rejected withResponse rejected (OutputLimit); keeping original transcription. Requests now send an explicit cap. - Completed transcriptions can no longer be overwritten: a failure writer could replace an already-saved
Completedresult. - Vocabulary replacements:
ApplyReplacementsnow trims only when vocabulary is actually configured. - Paragraph breaks: the local API
/post-processendpoint dropped dictated paragraph breaks. - Backup export: fixed a crash (
NullReferenceException) in backup export settings. - Reduced Sentry noise from benign no-speech results.