3.10.0

No More Phantom Text

• On-device Whisper answers a silent stretch of audio with something it memorized during training rather than with nothing, so a line you never dictated could land in your transcript
• VivaDicta now reads the model's own confidence numbers to tell an invented line apart from quiet real speech, and drops it before it reaches your note
• English, Russian, Turkish, Czech, and Arabic are all covered
• The words alone are never matched, so a "Thank you." you actually dictated still comes through

Your Recording Survives

• A phone call, an alarm, or Siri used to stop the capture at the system level without telling the app, so the waveform flatlined and the minutes you had already spoken were gone
• An interruption is now treated as a Stop you never pressed, and whatever was recorded up to that point is transcribed and saved like any other note
• When transcription itself fails, the recording is kept as a note you can play back and retry instead of disappearing with the error
• A recording that fails to start now tells you why, with a shortcut to Settings when the microphone was denied

Your Music Keeps Playing

• Opening a note used to pause whatever you had going in Spotify, YouTube, or a podcast app, without a single tap on play
• VivaDicta now takes over the audio only when you actually start playback
• It hands the audio back when you stop or close the note, so the other app resumes

The Newest Models

• GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, and GLM-5.3 join AI processing
• Gemini 3.8 Flash becomes the new Gemini default for both transcription and processing
• Cartesia Ink-2 and Gemini 3.5 Transcribe Live join the real-time models that transcribe while you talk

Czech Keyboard

• A proper Czech QWERTZ layout, with y and z where Czech expects them
• Callouts carry only Czech letters instead of a pan-European mix, so a letter with no Czech diacritic gets no callout at all
• A language added in a later update now switches itself on if you already use it in iOS, rather than waiting to be found in Settings

Back to 30 More Apps

• The keyboard's return list grew by thirty apps, VK Messenger among them
• Dictating from one of them sends you straight back instead of showing the manual return prompt

A Calmer Settings Screen

• The Transcription and AI sections had grown into walls of toggles, so the rows only a few people touch moved one level down
• The Gemini thinking level and transcription instruction now live in Advanced, and the two extraction toggles sit together behind a single Reminders and Calendar row
• The Live Translation button in the toolbar is off by default on fresh installs. If you already have it, it stays

Fixes

• A downloaded on-device AI model is recognized as ready everywhere in the app, so the mode editor no longer asks you to download a model you already have
• OpenCode requests now identify the app properly, which lifts the rate limits an anonymous client was getting and keeps the prompt cache warm between calls