Whistle offers local speech recognition in a 16.9 MB model

Whistle is a speech-recognition model that runs on a CPU through Cactus’s existing engine. It accepts up to thirty seconds of audio per pass in seven supported languages and can return word timestamps or speech embeddings. Loading it alongside Needle lets an application feed the transcript into a separate tool-selection step within the same engine.

Cactus · Oct 2, 2026

Why it matters

Useful for evaluating offline voice interfaces on devices where sending audio to a hosted service is undesirable.

More context

Try the published weights and Python package or supported platform binary on your own recordings. The base input is 16 kHz mono audio; microphone capture and resampling require additional dependencies. Keyword biasing can help names, but supported languages do not include every language.

Keep in mind

The 16.9 MB figure is the model file size, not total runtime memory. Reported speed uses an Apple M4 Pro and whole audio clips, not an arbitrary microcontroller or streaming setup. Benchmark accuracy comparisons mix published results; we have not tested them.

Original read