Open-Source Guide: Running a Fully Offline Voice Agent on Android in 1.2GB RAM
ivan_digital · reddit · 2026-08-13
A developer successfully deployed a fully offline voice assistant on a Galaxy S23 Ultra. The complete voice-to-action pipeline uses only about 1.2GB of memory, running speech recognition, action selection, and synthesis locally without internet.
Tech Stack & Highlights:
- Pipeline: Silero VAD → Parakeet-EOU 120M STT → FunctionGemma 270M → Android action → Pocket TTS.
- Model Control: FunctionGemma uses a 9.5MB adapter to output structured tool calls rather than open-ended responses.
- State Filtering: Filters available tools based on current device state (e.g., exposing "stop music" only when music is playing) to prevent invalid actions.
- Safety Checks: The model is not allowed to invent outcomes; contact and music searches are verified against real device data.
The prototype currently supports contacts, dialing, local music, and volume control. The author discusses the trade-off between extending a constrained action router and introducing a larger planner.
More from coding & agent
- Devs say LLM coding mistakes are a different class: overcomplication, inferred limits, worse at higher reasoning — Mu5_ · 2026-10-16
- How do you debug bad agent runs when you already have tracing? — Sensitive-Parsnip-12 · 2026-10-16
- Where does pre-task context live for agents? Devs debate workflows beyond handoff — shishen2026 · 2026-10-16
- No-instruction autonomous agent wakes up, sets own goals, and tries hacking by default — sethlazar · 2026-10-16
- CODESTRUCT reframes codebases as AST action spaces to fix brittle LLM code agents — burkov · 2026-10-16
- Would you install another SDK for agent debugging on top of your tracing stack? — Sensitive-Parsnip-12 · 2026-10-16