Box

[jegly]

v1.0.2 — Voice Mode & Vision Talk / NPU/TPU Acceleration

New Features

Free Talk — Hands-Free Voice Conversation

• Continuous speech-to-speech AI conversation with no button press between turns
• Tap the mic button to start, speak naturally, AI replies and listens again automatically
• Real-time TTS: each sentence is spoken as it generates rather than waiting for the full response
• Enable Real-time voice reply in Settings to turn it on

Vision Talk — Live Camera + Voice AI

• Stream your back camera live to the AI during a voice conversation
• Point at anything and ask about it out loud — AI sees the current frame and speaks the answer back
• Fully hands-free: mic → AI sees scene → AI speaks → mic restarts automatically
• Requires a vision-capable model (Gemma 4 E2B or E4B)

TTS Voice Picker

• Choose your preferred offline system voice in Settings
• Only locally installed voices shown (network-dependent voices filtered out)
• GrapheneOS / de-Googled ROM users: install RHVoice or eSpeak NG from F-Droid

LaTeX Math Rendering

• Greek letters, operators, fractions, integrals, summations rendered inline as Unicode
• Works automatically in assistant messages

Improvements

• GGUF model context size now scales with file size to prevent hangs on Android 17 Beta 4 (MemoryLimiter) — large models use a reduced KV cache automatically
• Removed QNN Hexagon skeleton .so libraries ( will be added back next release to support snapdragon acceleration )
• Removed interactive map skill

Bug Fixes

• Fixed model initialization hanging indefinitely — now times out after 30 seconds with a clear error
• Fixed app crash on invalid task ID (force unwrap replaced with safe exit)
• Fixed real-time TTS stalling on responses with no sentence punctuation (code blocks, URLs now flush correctly)
• Fixed TTS audio playback wiring in AI Chat

---

This discussion was created from the release Box .