Moonshine AI has released an open-source AI toolkit called Moonshine Voice, designed for developers building real-time voice applications that run entirely on-device. The framework emphasizes speed, privacy, and ease of use, requiring no internet connection, API keys, or accounts. It is optimized for live streaming scenarios, delivering low-latency responses by processing speech incrementally as users talk. Moonshine Voice supports a wide range of platforms including Python, iOS, Android, macOS, Linux, Windows, Raspberry Pi, and IoT devices. The library includes high-level APIs for common tasks such as speech-to-text transcription, speaker diarization, and command recognition, making it accessible even to developers without deep expertise in speech technology. The toolkit features a family of custom-built speech recognition models trained from scratch, offering better performance than Whisper Large V3 in live speech applications. These models include variants ranging from tiny 26MB versions for resource-constrained devices to larger models with up to 245 million parameters. Notably, the Moonshine Medium Streaming model achieves a lower word error rate than Whisper Large V3 despite using significantly fewer parameters—250 million versus 1.5 billion—making it more efficient for edge deployment. Key advantages over Whisper include flexible input windows (no fixed 30-second limit), caching of audio encodings to reduce redundant computation, and language-specific models that improve accuracy for non-English languages like Japanese, Korean, Arabic, and Mandarin. The system also supports real-time feedback during speech, enabling responsive user interfaces. The underlying architecture is built around a portable C++ core using ONNX Runtime for cross-platform performance, with native bindings available for Python, Swift, Java, and C++. This allows developers to use a single API across all major platforms. Moonshine Voice includes tools for debugging, such as saving raw audio input and logging API calls. It also provides a benchmarking tool to measure latency and compute load, with results showing that Moonshine models can be up to 5x faster than Whisper in live scenarios. The project is open-source under the MIT License for English models and the Moonshine Community License for others. The team welcomes contributions and offers support through Discord and GitHub. Future updates are expected to include advanced intent recognition with slot filling and support for additional languages. Developers can get started quickly using pre-built examples for each platform or via pip install moonshine-voice. The library is ideal for applications requiring low-latency, on-device voice interaction, such as smart home controls, wearables, robotics, and real-time transcription tools.