Soniqo's speech-core v0.0.14 adds Smart Turn v3.2 end-of-turn gate before LLM invocation in local voice pipeline

Developers seek ultra-low-latency alternatives to Silero VAD for speculative voice-agent pipelines

An open-source voice agent that runs entirely on an Android phone using a 270M parameter LLM (FunctionGemma) with LoRA adapters for tool calling. It performs VAD, streaming STT, tool-calling, device actions, and TTS in 1.2 GB RAM, responding 0.9s after user stops speaking. Solves on-device agent reliability for mobile users without cloud dependency.

Products·building

Audio perception layer for LLM agents with growing memory

An open-source framework that gives LLM agents continuous audio awareness beyond speech-to-text, detecting environmental sounds like birds, breaking glass, or smoke alarms. It runs locally with CLAP, Whisper, Silero VAD, and sqlite-vec, and uses fast-agent to support any LLM provider including local Ollama models.

An open-source framework that gives LLM agents continuous audio awareness beyond speech-to-text, detecting environmental sounds like birds, breaking glass, or smoke alarms. It runs locally with CLAP, Whisper, Silero VAD, and sqlite-vec, and uses fast-agent to support any LLM provider including local Ollama models.

A voice AI tutor that uses a small, fast model instead of a frontier model to achieve low latency. The scaffolding around the model provides intelligence, while the model choice prioritizes speed to first token. It solves the problem of voice agents feeling broken due to pauses, targeting developers building real-time voice applications.

Friendo is a call-native voice agent that can initiate phone calls or take live calls via LiveKit WebRTC. It uses a cascaded stack of Deepgram, Claude, and ElevenLabs to achieve low-latency voice interactions with full control over tool calls and memory. It solves the problem of robotic, laggy voice AI by optimizing latency and state management.