moonshine-ai/moonshine

Moonshine Voice is an open source toolkit for building voice applications like transcription tools, text to speech readers, and conversational voice agents. Everything runs directly on the device, meaning no internet connection, account, credit card, or API key is needed to use it. Because of this, it is fast and keeps audio private to the user's own machine. The project includes speech to text models the team trained from scratch, which the README says beat OpenAI's Whisper Large V3 on accuracy while running far faster, since Whisper needs seconds per response and Moonshine's smallest models respond in well under a second. There is also a tiny 1MB model version for devices with very limited memory or processing power. On top of transcription, the toolkit supports text to speech, voice cloning, identifying different speakers in an audio recording, and recognizing spoken commands, such as understanding a phrase like "turn on the lights" even if it is worded differently each time. The same library works across many platforms: Python, iOS, Android, macOS, Windows, Linux, Raspberry Pi, and even microcontrollers and small embedded chips. Setup differs by platform. Python users can install it with pip and run a single command to start listening to a microphone. Mobile and desktop users download prebuilt example apps from the project's GitHub releases page and open them in Xcode, Android Studio, or Visual Studio. Linux users can download and build a portable example with a couple of shell commands. Speech to text works in several languages including English, Spanish, Mandarin, Japanese, Korean, Vietnamese, Ukrainian, and Arabic. Text to speech supports even more languages. The README recommends Moonshine specifically for situations involving live, real-time speech, where waiting for a large cloud based model to respond would feel too slow, such as voice assistants or hands free interfaces. The project also has a Discord community for support and video walkthroughs linked from the README to help new users get started. The full README is longer than what was shown.