Описание
Moonshine — компактная модель распознавания речи: пересчёт внимания вместо кэширования ключей — в 5 раз быстрее реального времени на CPU, точность уровня Whisper large при куда меньшем размере, запуск на ноутбуках, Raspberry Pi и в браузере, открытые веса и привязки к Python.
Что умеет:
— распознавать речь быстрее реального времени;
— работать на CPU;
— запускаться на встроенных устройствах;
— обеспечивать точность уровня Whisper large.
Полная документация ниже — автоматически из репозитория usefulsensors/moonshine (ветка main). Оригинальные материалы проекта распространяются на условиях его лицензии.
Moonshine Voice
Voice Interfaces for Everyone
Moonshine Voice is an open source AI toolkit for developers building real-time voice agents and applications.
Full guides, models, and API reference are at moonshine-voice.readthedocs.io.
- Everything runs on-device — fast, private, and with no account or API keys.
- Optimized for live streaming, with low latency by doing work while the user is still talking.
- Speech to text models trained from scratch, from higher accuracy than Whisper Large V3 down to tiny 1MB models.
- One library across Python, JavaScript/WASM, iOS, Android, macOS, Linux, Windows, and Raspberry Pi.
Quickstart
pip install moonshine-voice
moonshine-voice mic --language en
Every other platform is covered in the Quickstart, with runnable samples in Examples.
Building with a coding agent? Copy .agents/skills/moonshine-voice/ into your project’s .agents/skills/ folder, or run npx skills add moonshine-ai/moonshine --skill moonshine-voice. The skill teaches the current API shape so the agent does not reach for Whisper or the old DialogFlow names.
Documentation
- Quickstart to install and run on your platform.
- Using the Library for transcription, text to speech, and conversational agents.
- Models for what’s available, accuracy, and domain customization.
- API Reference for the classes, options, and C API.
Support
- Discord for live support.
- GitHub Issues for bugs and feature requests.
License
Licensed under the MIT License. The models are MIT by default too, in every language and at every size — the only exceptions are the legacy non-streaming models for languages other than English, which stay under the non-commercial Moonshine Community License and are enumerated in LICENSE.