Whistle debuts as a 16.9MB on-device speech recognition model for CPUs, supporting seven languages and fast tool-call workflows

Whistle is a new speech-to-text model designed for phones, wearables, robots, smart home devices, cars, and microcontrollers. The 16.9MB model runs locally on CPUs with no dependencies, supports seven languages, provides word timestamps and speech embeddings, and can load beside Needle to turn spoken clips into tool calls.

Whistle reflects the push toward edge AI models that prioritize low latency, privacy, and small footprints over cloud-scale processing. By sharing an engine with Needle, it points to a broader trend of combining local transcription with on-device intent handling for embedded and ambient computing systems.