Cactus Compute released Whistle, an open speech recognition model. It is a 16.9 MB file. It runs on CPU with no dependencies. It shares an engine with Needle. It transcribes seven languages. It reaches the first token in 11 ms. It can turn a clip straight into tool calls. It supports word timestamps and speech embeddings. It detects language automatically unless specified. It runs in a browser sandbox. It outperforms Whisper base and Moonshine tiny v2 on most benchmarks. It deploys to seventeen targets including mobile, wearables and microcontrollers.
What changed
Speech recognition models were larger, slower to start, and required separate engines or dependencies.
What it unlocks
Running speech recognition and tool-calling from a single binary on-device with no external dependencies.
- 16.9 MB model size
- 11 ms time to first token
- 1,319 tokens per second decode speed
- 7 supported languages
What you need to act on it
- a 16 kHz mono audio clip
- Python or C++ environment with Cactus Needle installed
Sources