Whistle: Speech to Text in 16.9 MB
Summary
Cactus Compute released Whistle, a 16.9 MB on-device speech-to-text model. It runs on CPU without dependencies, supports seven languages, and integrates with Needle to enable tool calls directly from audio. The article benchmarks Whistle against Whisper and Moonshine, emphasizing on-device privacy and deployment across mobile, wearables, and embedded devices.