v3.1.2
AudarAI/Audar-ASR-V1v3.1.2Aug 23, 2026by github-actions[bot]
AI Summary
This release addresses stability issues by fixing settings persistence, improving error reporting, optimizing memory usage, and correcting audio recording levels.
Key Highlights
- Settings now persist correctly after restart.
- Detailed error reporting with exit codes and signals instead of generic messages.
- Reduced memory footprint for model loading.
- Corrected audio recording levels (9 dB louder).
New Features
- Persistent settings after restart
- Detailed error reporting (exit codes and signals)
- Broken model download detection and re-download prompt
- Reduced model loading memory usage
- CPU AVX2 naming support
- Functional phone pairing QR code
- Louder audio recordings
- Two-key dictation shortcut
- Dismissible yt-dlp update prompt
- Agent skill installation in Claude Code
- Windows VC++ runtime version checking
Full Release Notes
What's new? 🎉📣 - ⚙️ **Your settings stick again** — every setting reverted to its default on restart in 3.1.1. The file on disk was always correct; the app simply never read it back. If you gave up re-setting your preferences, they are safe now - 🔍 **Failures say why** — a sidecar that dies now reports its exit code and signal instead of a bare "sona process died", and a crash mid-transcription reads as a crash rather than "error decoding response body". Whisper's own diagnostics reach the log again, after being silently discarded by a no-op callback - 📦 **Broken model downloads are caught** — files are verified before they are used, incomplete ones are detected on startup and offered a re-download, and an interrupted download can no longer masquerade as an installed model - 🧠 **Half the memory to load a model** — measured 294 MB → 216 MB on a small model, and roughly one model copy less on large ones. Machines that used to run out of memory mid-load have room now - 🖥️ **A CPU without AVX2 is named** — instead of dying anonymously, Vibe says the processor is unsupported and why - 📱 **The phone pairing QR scans** — one wrong character in the QR generator misplaced the alignment patterns on any code past a certain size, and the pairing code was just past it. No scanner could read it - 🎙️ **Recordings are 9 dB louder** — mixing in system audio was quietly halving the level, then halving it again. Every recording made with system audio selected was far quieter than it should have been - ⌨️ **Dictation defaults to a two-key shortcut** — Option+Space on macOS, Ctrl+Space elsewhere, instead of a three-key chord - 🔗 **yt-dlp stops nagging** — the update prompt appeared on every launch and forgot "Later" each time. It is now a dismissible toast at most once a week, and it offers the update when a download actually fails, which is when it helps - 🤖 **Install the agent skill** — `Settings → API & Agents` writes it to Claude Code and Codex so it is available in every session, with the transcripts folder and the local API baked in - 🪟 **Windows: the VC++ runtime is version-checked** — the installer only asked whether *any* runtime was present, so machines with an old one were skipped and crashed later with no message - 🌍 The product name and a mistranslated setting fixed across all 22 languages Signed and notarized on macOS and Windows. **Full Changelog**: https://github.com/thewh1teagle/vibe/compare/v3.1.1...v3.1.2