# Speech Core

An on-device speech infrastructure library in C++17 for Linux, Windows and Android, providing voice activity detection, streaming speech-to-text, speaker diarization, text-to-speech and a voice-agent pipeline.

Speech Core runs locally on CPU with no cloud calls, no Python needed at inference time, and no audio leaving the machine. It separates a small, model-agnostic orchestration layer, handling turn detection, interruption, audio utilities, conversation state and tool calls, from optional ONNX Runtime or LiteRT inference backends, so an application chooses its own models rather than being locked to one.

It is built for live conversational agents, with voice-activity-driven turn taking, eager streaming transcription that produces partial results as a user speaks, barge-in handling, and streaming text-to-speech. Recent additions include an OpenAI-compatible local text-to-speech server endpoint, a self-contained Windows release package bundling the server and its models, DeepFilterNet3-compatible noise suppression, and a fixed-frame streaming text-to-speech backend. It exposes both a native C++ API and C bindings suitable for Kotlin, JNI or Swift hosts, and is licensed Apache-2.0.

Website: https://soniqo.audio/speech-core