Loquara 0.6.4

Loquara 0.6.4 brings cloud transcription and xAI live typing to the latest public release, with fixes for the recording pill and repeated text. The existing interface is retained.

Fixes

• Restore the microphone animation when recording starts after the overlay has loaded.
• Prevent duplicate live text from oversized transcript revisions and automatic fallback after uncertain input delivery.
• Preserve spacing between live segments and stop replacement typing when a correction fails.
• Stop cancelled or timed-out live writers before they can send more input, and drain queued audio before finalizing the stream.
• Show delivery failures on the overlay and restrict provider model lists to speech models.

Cloud dictation

• Optional API-key transcription with xAI, OpenAI, Groq, Mistral, ElevenLabs, Deepgram, and OpenAI-compatible endpoints.
• xAI streaming types words while you speak. Uploaded recordings use 16 kHz mono audio.
• API keys are encrypted with Windows DPAPI. Local transcription remains available.

Known limitations

• Keep the same text field focused during live dictation; input and corrections do not yet guard against focus changes.
• An interrupted live stream can leave a partial transcript marked as completed in History.
• Rapid cancellation and retry of the same saved recording can race an earlier batch transcription.

Notes

Windows x64 installers are provided as EXE and MSI. Install over the existing version; settings and recordings in AppData are preserved.

Validation: 108 frontend tests and 207 Rust tests passed. Two environment-dependent Rust tests were skipped (downloaded-model transcription and a real-host download).

Loquara 0.4.4

History now keeps a recording available when transcription did not complete. A saved WAV can be transcribed again from the recording inspector.

History

Failed or cancelled recordings with saved audio now show Retry. Completed recordings with an empty transcript are labeled No transcript and can be retried. The action reports when the audio file is no longer available.

Notes

Install over 0.4.3. Recordings in AppData are left in place. Select a failed or empty-transcript recording in History and choose Retry.

Loquara 0.4.3

The dictation pill stays above the window you are speaking into. It still does not take the keyboard.

Overlay

Showing the overlay without activating it left the pill under the focused app. Each show now pins the window as topmost without stealing focus.

A remembered position that sits off every display is pulled back onto the virtual desktop. A pill on a second monitor stays there.

Notes

Install over 0.4.2. Settings → Behavior → Show the overlay must remain on.

Loquara 0.4.2

Long recordings no longer abort the app when you press Retry. A stalled microphone queue no longer deletes the take.

Transcription

Parakeet used to see the whole file at once. Several minutes of audio made the native encoder abort (Windows error 0xc0000409). Takes are now transcribed in 30-second windows and joined. A seven-minute recording that crashed 0.4.0 completes.

Recording

A full microphone queue used to fail Stop and delete the WAV. The take is kept. Cancelled long recordings keep audio for Undo. History no longer claims a file exists when it does not.

Capture duration is no longer overwritten with how long transcription took. Overlay and History errors for these cases are shown in Polish.

Notes

Install over 0.4.0. Recordings in AppData are left in place. Retry a failed long take from History after installing.

Loquara 0.4.0

The compact pill stays the default. A larger recording window is optional. Cancelling no longer steals the window you were dictating into. Opening a transcript from the home list is immediate.

Overlay sizes

Settings → Behavior → Overlay size:

• Compact (default): the 68×36 pill. Hover shows Stop and Cancel.
• Expanded: a larger recording window with waveform, elapsed time, mode, Stop and Cancel.

Right-click the overlay to switch without opening Settings.

The meter draws a short history of loudness rather than a lockstep equalizer. The expanded window uses a longer window and a slower motion while processing.

Cancelling

Escape or × discards a short take immediately. Longer recordings show Undo for five seconds on the overlay itself — no card, and the overlay does not take focus from the application you were dictating into. Processing can be cancelled from the overlay.

The previous confirm card (Escape to ask, Enter to discard, ten-second timeout) is gone.

History

Clicking a recent transcript on the home screen opens that recording. Audio is fetched only when you press Play. Waveforms off-screen wait until they are near view, so the first paint no longer waits on every canvas and a WAV decode.

Notes

Install over 0.3.2. Compact remains the default overlay size.

Loquara 0.3.2

Reliable pasting into SSH sessions, terminals and Electron applications, along with a fix for window sizing when inserting text.

Pasting into terminals and SSH clients

Injecting dictation into terminals (such as Termius, Windows Terminal, PuTTY/KiTTY, Git Bash, Alacritty, WezTerm, WSL) now works automatically:

• Hardware scan codes: Keystrokes sent via SendInput now populate virtual scan codes (wScan), allowing Electron/Chromium-based terminal emulators (e.g. Termius / xterm.js) to recognize synthetic paste inputs.
• Key hold duration: Modifier keys are held down briefly across message loops to ensure combinations like Ctrl+Shift+V are registered reliably.
• Focus synchronization: The application waits for the target terminal window and its renderer process to activate before firing keystrokes.
• Smart shortcut selection: Terminals that expect Ctrl+Shift+V or Shift+Insert are detected automatically by process name and window title.

Configurable Paste Mode

A new Paste mode option in Settings → Behavior allows choosing how text is pasted:

• Automatic (recommended): Auto-selects Ctrl+V, Ctrl+Shift+V, or Shift+Insert depending on the active target window.
• Terminal / SSH (Ctrl + Shift + V): Forces the Linux terminal paste shortcut.
• Universal terminal (Shift + Insert): Alternative shortcut for PuTTY and older terminal emulators.
• Standard (Ctrl + V): Standard clipboard paste.

Fixes

• Maximized & split-screen window resizing: Pasting into a snapped or maximized window no longer restores it to an unsnapped, centered rectangle.
• Restoring foreground focus now attaches thread input queues, avoiding Windows foreground activation lock refusals.

Loquara 0.3.1

A follow-up to 0.3.0, clearing out what the move away from Python left behind. Everything in 0.3.0 applies.

A setting that did nothing is gone

Settings offered a "Dictation language" choice, described as forcing the language "instead of auto-detection (Whisper/Cohere)". Neither model has been in Loquara since 0.3.0, and the setting had stopped doing anything at all: the value was saved and never reached the engine. Parakeet detects the language on its own and takes no hint, so the control has been removed rather than relabelled.

Text from a version that no longer exists

The "Before you start" screen walked through installing Python 3.10, engine dependencies and a PyTorch build matched to your graphics card. None of that has been needed since 0.3.0. Those strings, and the errors pointing at a missing Python, have been removed.

Builds

The build now stops if any library the engine needs is missing, instead of producing an installer that starts and then fails on the first dictation — which is how 0.2.0 shipped without them. Continuous integration builds and tests on Windows again; the Linux job was removed, since the engine links DirectML and the platform layer is Win32.

Loquara 0.3.0

Loquara no longer needs Python. Transcription runs inside the application through ONNX Runtime, so there is no interpreter, no PyTorch and no CUDA toolkit to install first. The installer carries everything except the model.

Runs on any graphics card, and on none

The application does not inspect your hardware. It starts DirectML and falls back to the processor if that fails, which is one path covering NVIDIA, AMD, Intel and a machine with no graphics card at all. Measured on 30 seconds of audio: 0.32 s on the card, 0.88 s on the processor.

The model

The one thing the installer does not carry, because it weighs 670 MB. On first run the dictation screen offers to fetch it — no account, no token, no licence to accept.

• Four connections at once. Measured: 3.3 MB/s over one, 13.3 MB/s over four.
• Progress shows on the dictation screen, in Settings and on the badge in the corner. The transfer belongs to the application, so moving between screens neither loses it nor starts a second one.
• The download can be stopped.
• Not enough disk space is reported before the transfer rather than 600 MB into it.
• Files left by an interrupted download are cleared on the next start.

Discarding a recording

Escape asks, Escape again takes the question back, Enter discards. Previously one key both asked and confirmed, so two accidental presses threw a recording away.

The question appears as a card beside the pill instead of stretching it. When the pill sits at the top of the screen, the card opens below it.

The recorder pill

Shorter and narrower, 68 by 36. The meter learns the room's own level and draws only what rises above it, in a colour that separates a voice from a fan — it used to meter raw loudness and sat half-height whether or not anyone was speaking.

Starting up

A new "Start in the tray" setting: no window on launch, and the shortcut works right away. It pairs with launching at login.

Fixes

• The window stopped responding while a model loaded, because asking the engine for its state waited on the same lock as the load.
• The panel could claim there was no model while one sat on disk and dictation worked.
• Pressing Escape mid-recording froze the application.
• Escape was held globally for the application's whole life, so it stopped working in other programs.
• "Open model folder" and revealing a recording opened Documents instead of the right folder.
• Toggling one setting made the other controls flicker.

Loquara v0.2.0

Dark-only redesign, real waveforms, and two fixes worth upgrading for.

Fixed

• Escape no longer breaks in other applications. 0.1.0 registered Escape as a global hotkey at startup and held it for the whole session, so no other program ever saw the key — games, dialogs and menus included. It is now claimed only while a recording is in progress.
• Cancelling a recording works. The confirmation could be raised but never answered, because its keys lived in a window that is shown without focus. Press Escape once to ask, again to confirm; wait ten seconds and it withdraws.

Changed

• Dark-only interface with a rebuilt type and colour system. The light theme and its setting are gone.
• Recordings store a real amplitude envelope, captured as the audio is written. History and the player draw that instead of a shape derived from the recording id, and the player uses it as its scrub track.
• The recorder pill is 80px and widens only when it has something to say.
• Gated Hugging Face models now explain the account, licence and token steps in place. Tokens are verified before they are stored, and kept in the Windows Credential Manager rather than the app database.
• Dropdowns are drawn by the app instead of the operating system.
• Changing one setting no longer greys out every other control.

Notes

Existing recordings keep working; the database migrates itself. They show no waveform, because no envelope was stored when they were made.

The Linux packages are produced by CI but have not been tested. Windows is the supported platform.

Loquara v0.1.0

Local, offline dictation for Windows (Whisper / Parakeet / Cohere). Everything runs on your machine — audio never leaves it.

Installer (Windows x64): Loquara_0.1.0_x64-setup.exe

Requirements: Windows 10/11 x64, Python 3.13 + PyTorch (CUDA for GPU). Models are downloaded from the in-app Settings screen (Whisper Large v3 Turbo, Parakeet TDT, Cohere Transcribe).

Key features

• Dictate with a global shortcut (Ctrl+Space), Superwhisper-style floating pill
• Fully local transcription — no cloud
• Long recordings (>30 s) now transcribe correctly with Whisper
• In-app audio playback with seek + reveal file location
• Dictation language setting (Auto / PL / EN) and model keep-alive setting
• Custom titlebar (drag anywhere, minimize/maximize/close)
• History, transcript export, copy recording path, clear failed recordings