FluidVoice Windows 0.0.12

• File Transcription: drop an audio or video file (MP3, M4A, WAV, FLAC, OGG, Opus, MP4, MOV, MKV, WebM and more) on the new File Transcription page to get a transcript with your active voice model. Recordings up to 3 hours are supported (6 on a graphics card). Transcripts are kept in a panel on the right where you can rename, copy, export or delete them. Dictation pauses while a file is being transcribed.
• Local API (Settings > Experimental, off by default): lets scripts on this PC send files to FluidVoice for transcription, using a token.
• Settings saved with a byte-order mark (for example by Notepad) are no longer treated as damaged and reset.
• Smart capitalization is now on and reads the text around your cursor. Dictating into the middle of a sentence starts lowercase, leaves off the final period and adds any missing space; after a sentence or in an empty box it starts with a capital. Search and address boxes and short replies like "Sounds good" no longer get a period. Where the text can't be read, your dictation is left as it is.
• If a dictation doesn't land in the text box, FluidVoice pastes it once more. If that also misses, the text stays on your clipboard and a small card says to press Ctrl+V. You can turn this off from the card or in Settings > Text.
• AI enhancement now runs on the graphics card for older NVIDIA cards too (GTX 900 and 1000 series, older Quadros, Titan V) with a driver from 2024 or later (#988).
• AI model downloads pick up where they stopped after a dropped or stalled connection, instead of hanging at a random percentage (#1140).
• On graphics cards that need a one-time setup, a "Setting up your GPU" card shows during the first, slower load of the speech model (#1177).
• The recording overlay and its cards fit when Windows Text size is above 100%, and stay above a taller-than-usual taskbar (#1176).
• If the Start with Windows entry points at another program, FluidVoice fixes it on its next start (#985).
• A short, optional survey after the welcome screen, shown once.
• A cleaner, calmer look. FluidVoice now draws its own title bar with a button to collapse the sidebar to icons, uses a black, white and grey theme, and has a new Dashboard with your stats, recent dictations and shortcut. Pages stay centred and use the extra room when the window is maximised.
• New Appearance settings (Settings > General): pick an accent colour and white or cream text. Ice blue is the new default, and Insights and setup pick up your accent.
• Recent dictations and History show the logo of the app you dictated into.
• The Voice Engine page is simpler: Fluid Speech Mini is listed first as recommended, and each model shows one line plus speed and accuracy bars.
• On PCs where speech runs on the processor, live preview is now off by default and, when switched on, uses far less CPU while you dictate.
• You can remove the Esc shortcut that cancels a dictation, so Esc reaches the app you are working in. Canceled dictations are saved to History.
• Added Fluid Speech Mini, our own speech model for the same 25 languages as Parakeet v3, including English, German and Spanish. It is about a third of the download of the other voice models (331 MB), uses about half the memory, and holds up better with someone else talking nearby. Setup now suggests it first for those languages.
• Added Fluid-1.1 Pico (research preview), our smallest AI enhancement model: 697 MB and about twice as fast as Fluid-1 Mini without a graphics card. Setup now downloads it in the background while you finish the other steps, so you can try AI polishing straight away.
• Cleanup Style for Fluid-1.1 Pico: Standard (balanced), Light (keeps your wording and adds no lists) or More paragraphs.
• The AI Enhancements page shows the models side by side. Activate a model, or Deactivate it to turn AI off and free its memory, and FluidVoice recommends a model based on your PC's memory and graphics card.
• Fluid-1 Mini is no longer marked as a preview. If you already use Fluid-1 or Fluid-1 Mini nothing changes, and every model is still available in AI Enhancements.
• Setup lets you choose Toggle, Hold or Double Tap for the dictation shortcut, and the practice step is clearer.
• Download progress bars now move smoothly instead of stalling and jumping to done.
• If the graphics card fails while recognising speech, FluidVoice retries that dictation on the processor instead of losing it.
• The speech engine and the AI engine now choose their graphics runtime separately, so an NVIDIA driver too old for one no longer holds the other back.
• Fixed dictated text staying on the clipboard instead of your previous clipboard coming back. It happened in Google Docs-style editors, after apps that rewrite pasted text, when the earlier clipboard held an image or files, and in every app when a clipboard sync or remote-desktop tool was running. FluidVoice now waits until the app has actually taken the text, then always puts back what you had copied. If an app never takes the paste, it is typed instead. A slow app can no longer end up pasting your old clipboard.
• Pasting is more dependable. Your previous clipboard comes back even in slow apps and across back-to-back dictations, a dictation is never pasted twice, and the text is kept for recovery when delivery cannot be confirmed.
• Shortcut fixes: AltGr no longer sets off Ctrl+Alt shortcuts, keys released on the lock screen no longer stay stuck, and Escape cancels setting a shortcut instead of leaving the keyboard captured.
• Speech and AI models are released when idle according to available memory.
• The recording overlay stays on top, on the monitor you are working on.
• FluidVoice no longer flashes its window when it starts with Windows or restarts itself while idle.
• Exported logs no longer contain dictation identifiers or timestamps.