Moonshine

Moonshine targets speech recognition on hardware where Whisper is impractical, from laptops down to Raspberry Pi boards, microcontrollers, and wearables. The models scale from 26M parameters up to a 245M medium streaming variant that reaches 6.65 percent word error rate, better than the 7.44 percent of Whisper large-v3, while transcribing a clip in 107 ms on a MacBook Pro versus more than 11 seconds for Whisper. Streaming variants cache encoder and decoder state to avoid redundant computation, and the repository has grown into the Moonshine Voice toolkit, adding on-device text-to-speech in 16 languages, voice activity detection, speaker identification, and intent recognition, with speech recognition covering English, Spanish, Mandarin, Japanese, Korean, Vietnamese, Ukrainian, and Arabic. Inference runs through ONNX Runtime; installation is pip install moonshine-voice, with Swift Package Manager, Maven Central, and C++ libraries covering iOS, Android, and Windows. Everything executes locally with no API keys, and the MIT license covers code and models, from Moonshine AI, formerly Useful Sensors.