Whistle: on-device speech to text in 16.9 MB
Whistle is an open speech recognition model from Cactus that fits in a single 16.9 MB file and runs on a CPU without dependencies. It transcribes seven languages with word timestamps, which is handy for captions, and audio never leaves the device.
What it does
Whistle is an open speech recognition model from Cactus in a single 16.9 MB file that runs on a CPU with no dependencies. It transcribes English, German, French, Spanish, Italian, Dutch and Polish, detects the language, and gives word-level timestamps.
It can also bias toward keywords you pass in, which helps with names and jargon.
Results and limits
Cactus reports 11 ms to the first token and compares speed against Whisper and Moonshine on an Apple M4 Pro CPU, all in its own tests. A transcript is capped at 320 tokens per window, and longer audio is processed in windows.
Worth knowing
Weights are on Hugging Face and source code on GitHub. There's an in-browser sandbox where audio never leaves your device. It's speech-to-text only, useful for captions and logging footage.
Try it yourself
Open the Whistle sandbox, upload [a 30-second interview clip], add keywords [character names, place names], and compare the word timestamps against your edit for rough captions.