Voclaude 0.5.0

Voclaude 0.5.0

Install: run voclaude-v0.5.0-gpu-setup.exe (per-user, no admin prompt, optional "start with Windows"). The zip still works if you prefer to unzip and run.

New

• One-click installer for Windows (Inno Setup): Start Menu entry, autostart option, uninstaller, app icon.
• Local API for scripts and agents. While the app runs it serves an OpenAI-compatible endpoint on http://127.0.0.1:7770/v1/audio/transcriptions that accepts uploads, local paths ({"path": "..."}) or raw bodies, decodes audio and video files (wav, mp3, flac, ogg, m4a/mp4/mov, mkv; anything else through ffmpeg if installed) and uses the same loaded model as dictation. voclaude transcribe <file>... is the command-line client. Results appear in History. See docs/API.md.
• CPU-only machines are now usable. Q8 weights run on the CPU with AVX2 kernels: the 1.7B model at about 2.3x realtime and the 0.6B model at 3.4x on a desktop i5, up from 0.9x and 1.9x. Because segments are transcribed while you speak, this keeps up with dictation.
• Apple Silicon (experimental). A Metal device is created when built with --features metal; inference verified on GitHub's M1 runner (1.7B Q8: 4.6x realtime). The tray/hotkey UI does not run on macOS yet.

Fixed

• Clean installs failed on first run because the Qwen3-ASR repositories do not publish tokenizer.json; the tokenizer is now rebuilt from vocab.json, merges.txt and tokenizer_config.json (#1, thanks @oxymoron304).
• The GPU build now really targets compute capability 7.5 (GTX 16xx / RTX 20xx and newer). The bundled CUDA kernels previously only compiled for the build machine's own GPU generation.

Requirements

• Windows 10/11 64-bit, a CPU with AVX2 (any since 2013).
• NVIDIA GPU with 6 GB or more and a recent driver for the fast path (CUDA 13 runtime DLLs are bundled); otherwise the CPU path is used automatically.
• The model (about 4.7 GB) downloads from Hugging Face on first use.

The installer is not code-signed yet, so SmartScreen may ask for "More info" then "Run anyway".

Voclaude 0.4.0

Voclaude 0.4.0

Local voice-to-text for Windows: press F4, speak, press F4, paste.

Install: unzip, run voclaude.exe. First run downloads the model (~3.4 GB). Details and requirements in the README.

What's new since 0.3.1

• Transcribes while you speak: the recording is cut at pauses and transcribed in the background, so the text is on the clipboard 0.3–2.5 s after you stop (was 35 s for a 5-minute dictation).
• Never loses a word: audio synced to disk every 5 s, every finished segment journaled, automatic recovery after a crash or power cut, tray item "Recover Last Recording".
• 3x faster decoding (fused kernels + 8-bit weights), 3.6 GB VRAM instead of 5.9 GB.
• Idle footprint: inference runs in a child process that exits after 60 s idle — 0 VRAM, ~20 MB RAM.
• No NVIDIA GPU: falls back to the CPU with the 0.6B model automatically. Small GPUs also get the 0.6B model.
• Bundled CUDA runtime DLLs: nothing to install besides an NVIDIA driver from 2025 or newer.

Known limits: Windows only for now (a macOS build job exists but is unverified). CPU-only transcription is ~2x realtime, so long dictations finish later than on a GPU. Russian/English mixing is supported; the default model was chosen over Whisper large-v3-turbo because Whisper dropped Russian passages on our tests.

macOS (experimental): voclaude-v0.4.0-macos-experimental.zip is built by CI with Metal but has not been run on a Mac yet and is unsigned. To try it: unzip, then xattr -d com.apple.quarantine voclaude and run ./voclaude from a terminal; grant microphone and accessibility (global hotkey) permissions when asked. Please report what happens.