What's New
- Fluid-1 is 2.2x faster on Apple silicon. Existing users receive a one-time prompt to download and verify the new model, while keeping the previous model if they decline or the upgrade fails.
- FluidVoice can suggest Custom Dictionary replacements after you repeatedly correct dictated text.
- Train by Voice can save pronunciation samples for matching names and words that sound alike.
- Spoken punctuation formatting is now optional and configurable: choose the trigger phrase and edit its punctuation aliases in Custom Dictionary.
- Slash commands and @ phrases can be formatted while dictating, with the feature disabled by default.
- All Whisper models are back in the menu and it's faster than before!
- You can copy your most recent transcript from the menu bar, and click rows in Getting Started to open their relevant settings.
- AI dictation prompts can be configured to send only the dictated text instead of always adding FluidVoice's base prompt.
Improvements
- Fluid-1 verifies backend-specific downloads before switching and keeps the previous backend until the replacement is ready.
Fixes
- AI dictation now keeps each prompt's configured provider, fails closed when Fluid-1 is unavailable, preserves Local API processing with app-only routing, and clears retired Apple Intelligence settings without selecting a fallback.
- Voice-engine downloads now show continuous byte progress for Parakeet, Cohere, Nemotron, and Whisper models instead of appearing stuck at 0% or 3%.
- Custom Dictionary manual replacements can now use commas and other punctuation as the text FluidVoice should replace, and malformed duplicate entries no longer interrupt transcription.
- Train by Voice now preserves earlier pronunciation samples during retraining, includes them in app backups, and clears stale voice profiles after replacement edits or legacy backup restores.
- Custom Dictionary suggestions now respect Escape dismissals from any active app, and voice training continues correctly after an audio fallback.
- Custom punctuation aliases now drive symbol-context formatting consistently, including edited dictionary rules.
- Whisper readiness checks no longer create an empty model folder.
- Dictation AI enhancement now falls back to a standard response when a provider does not support streaming.
- Faster Recording Start now handles microphone layout changes correctly, avoids false fallbacks when audio processing is delayed, and uses the compatibility audio path if malformed audio is detected.
Contributors
Thanks to @matthewrball for Copy Last Transcript, @snehkagrana for clickable Getting Started rows, @mvanhorn for dictation-only AI prompts, @glinhard for Anthropic temperature handling, and @postoso for stronger Whisper test isolation in 1.6.3.