Changelog

• New offline local ASR model: Fun-ASR-Nano 2512 recognizes Chinese, English, Japanese, Cantonese, and Chinese dialects/accents with automatic language detection — no Python, CUDA, or separate server required. Choose the Q8_0 build for higher accuracy or the lighter Q4_K_M build for faster startup.
• Fun-ASR-Nano streams recognized text live while you speak, so the overlay shows words as they are decoded instead of only at the end.
• The dictation overlay is smoother: the thinking glow now pulses with your microphone level, the waveform bloom is visible from the first listening frame, and the capture pill is more compact.
• Transcription text stays on a single line with the newest words always visible, and the panel keeps balanced left/right spacing as text grows and shrinks.
• Fixed the spinner being clipped and the “Transcribing…” label being cut off when switching from listening to processing.

• Dictation now pastes into the app you were using, matching the familiar clipboard paste flow.
• Smoother live transcription and result panels: multi-line sizing, less jitter, and auto-dismiss after copy.
• More reliable AI automation console: stable confirm/progress UI and smoother compact↔expanded morphing.
• Crash reporting via Crashlytics (only when you allow analytics; Debug builds never upload).

• Added animated Liquid Gradient activity feedback while transcription and AI automation are running.
• Fixed animated gradient glow alignment and clipping in overlay panels.
• Reordered Settings so AI configuration appears before activity and history sections.

• Refined script automation: cleaner proposal handling, scaled overlay typography, and a tidier automation console.
• Centered voice-capture overlay text and upgraded the liquid-glass background effect.
• Extracted the realtime transcript accumulator for more stable live transcription updates.

• Fixed Release-build crashes caused by duplicate MLX class identities across ASR frameworks (Hub, Jinja, MLXNN, MLXLMCommon now share a single runtime copy).
• Prompt post-processing now automatically retries without a JSON schema when an OpenAI-compatible model rejects guided grammar or structured output, so transcription no longer fails.
• Fixed the language dropdown being clipped in the website navigation at tablet widths.
• Polished website translations (Japanese, Dutch, Traditional Chinese).

• Added Traditional Chinese, Japanese, German, Spanish, French, Norwegian, Dutch, and Portuguese translations.
• Improved speech-model downloads with fallback sources and live progress.
• Smoothed live transcription updates and fixed clipping and layout shifts in voice overlays.
• Fixed Chinese count ordinals from first through tenth.

• Fixed Siri Remote microphone audio in the release build. Buttons worked but voice capture did not, because notarization re-signing broke Apple’s PacketLogger. The release app now ships PacketLogger as a notarization-safe payload; Setup Assistant extracts it on install so voice capture works again.
• If you already installed the Packet Logger helper from an earlier build, open Setup Assistant and reinstall it once so the new PacketLogger runtime is extracted.

• Setup Assistant opens immediately on first launch so you can pair the Siri Remote and finish setup without an extra prompt in front of it.
• Fixed the Privacy Policy link in Settings → About (now opens https://couchvox.com/privacy/).
• Siri Remote icons use Apple’s system Siri symbol across the app for a cleaner, more consistent look.
• Website: redesigned pricing for Free / Basic / Pro / Lifetime plans, Polar product preselection on checkout, and expanded FAQ covering plans, BYOK, and Full Privacy.

• Full Privacy mode in General settings — keeps processing on your Mac by turning off analytics, remote ASR, remote LLM providers, and AI Automation (with a confirmation dialog).
• First-launch prompt for Events & Crash Reports, with an Events & Crash Reports toggle when Full Privacy is off.
• Review Events opens a public catalog of what analytics may send (no audio or transcript content).
• Privacy Policy updated for Full Privacy mode, consent, and the analytics events page.

• Redesigned installer DMG — drag CouchVox from the TV onto the sofa (Applications).
• Fixed microphone permission registration under Hardened Runtime (audio-input entitlement).
• Aligned the privileged helper identity for more reliable remote capture setup.
• Cleaned up the Setup Assistant incomplete-count badge in settings.

First test release of CouchVox.

What's included

• Siri Remote button mapping with per-app profiles.
• Floating virtual keyboard for across-the-room typing.
• Remote and local microphone dictation with transcription.
• AI post-processing and confirmation-based automation workflows.
• Sparkle auto-update support (this is the first versioned release).

Known limitations

• DMG is not notarized yet — Gatekeeper will show an "unidentified developer" prompt. Right-click → Open to bypass.
• Pro entitlements and licensing are not active in this build.

Requires macOS 26 or later, Apple silicon.