Cactus Compute releases 16.9 MB speech model for local CPU use

Cactus Compute released Whistle, a 16.9 MB speech recognition model that it says runs on a device's CPU without dependencies. It transcribes up to 30 seconds of audio in seven languages and can return word timestamps or speech representations without producing a transcript. The weights are on Hugging Face, with source and an engine on GitHub. Prebuilt engines cover seventeen targets, including Android, iOS, watchOS, Windows on ARM and RISC-V. Cactus reports better results than Whisper base on several speech benchmarks, while Whisper base leads on TED-LIUM, AMI and the MLS average. Whistle's transcript is capped at 320 tokens.