moonshine-ai/moonshine

Moonshine Voice is an open-source toolkit for developers building real-time voice agents and interfaces. It bundles streaming speech-to-text (STT), intent recognition, and text-to-speech (TTS) so an application can listen, decide, and answer without sending audio to a server. Its stated focus is voice interaction rather than a general-purpose model server.

What it does and how

Everything is designed to run on-device, with no account or API keys, and the pipeline is optimized for live streaming: it processes speech while the user is still talking instead of waiting for a complete recording Moonshine README. That is the mechanism behind the low-latency claim — partial audio is transcribed continuously, so downstream intent or response logic can begin earlier.

The speech models are trained from scratch and cover a size range from models the project says are more accurate than Whisper Large V3 down to roughly 1MB models for constrained devices; the sizes and accuracy are documented on the models page rather than in the README Moonshine README, model documentation.

Ways to use it

• Python quickstart: pip install moonshine-voice followed by the moonshine-voice mic command gives a local microphone demo; other platforms have their own instructions Quickstart.
• Multiple platforms from one library: Python, JavaScript/WASM, iOS, Android, macOS, Linux, Windows, and Raspberry Pi are listed as targets Moonshine README. Each path's exact prerequisites are on the platform-specific quickstart page platform setup.
• Inside a coding agent: the repo ships an agent skill, copyable into a project or installable via npx skills add, intended to teach an agent the current API instead of Whisper or old DialogFlow names agent skill.
• Full application surface: transcription, TTS, and conversational agents are covered in the usage guide, with runnable samples alongside usage guide, examples.

Requirements and constraints

• The README describes on-device execution as the default posture, so what you actually need follows from the platform you select; the docs are the authority on each target, not this summary.
• The core licensing is MIT, and models are MIT by default across languages and sizes. The exception is the legacy non-streaming models for languages other than English, which remain under the non-commercial Moonshine Community License and are listed in the license file license. Check that exception before shipping non-English legacy models commercially.
• The supplied evidence does not include independent latency or accuracy validation beyond the project's own links, and the HN comments are unverified user opinions rather than maintained-benchmark results.

Context

The project was announced as "Speech Recognition and TTS in less than 500kb" and drew a large discussion thread, but that thread is commentary, not evidence of production fitness HN thread.