All work
AI Agents & MCP

Local Voice Agent for Mac

A fully offline, real-time voice agent for Apple Silicon.

Mic to Silero VAD to Whisper (MLX) to a streaming Ollama LLM to per-sentence Kokoro TTS, with Hindi/English routing and barge-in.

What makes it serious
  • Fully offline pipeline on Apple Silicon
  • Silero VAD, Whisper MLX, Ollama, and Kokoro TTS
  • Hindi/English voice routing
  • Barge-in interruption and per-sentence streaming

A real-time, fully offline voice agent for Apple Silicon. The pipeline runs entirely on-device: microphone audio passes through Silero voice-activity detection, into Whisper (MLX) for transcription, then to a streaming Ollama LLM, and out through per-sentence Kokoro text-to-speech. It routes between Hindi and English, and supports barge-in so you can interrupt it mid-sentence like a real conversation.

Doing all of this locally means no cloud latency and no audio leaving the machine — private by construction. Streaming the LLM and speaking sentence by sentence keeps the response snappy rather than waiting for a full completion before talking. Running the whole stack on-device is what makes it usable for private conversations and reliable without a network connection, something cloud voice assistants cannot promise, and streaming the model keeps replies fast despite the local hardware.

Stack
PythonSilero VADWhisper MLXOllamaKokoro TTS
#voice-agent#offline#apple-silicon#whisper#ollama#tts