By No Code MBA · Facts checked
- Provider
- Cactus
- Announcement date
- Status at last check
- Generally available
- Model family
- Whistle
- Capabilities
- Audio, Embeddings
What changed
Whistle is an on-device speech recognition model packaged as a single 16.9 MB file that runs on CPU. It transcribes up to 30 seconds per pass in English, German, French, Spanish, Italian, Dutch and Polish, returns word-level timestamps, and can output speech embeddings at one row per 80 ms frame. Cactus reports 11.1 ms time to first token on a 10-second clip on an Apple M4 Pro CPU versus 73.2 ms for Whisper base, which is 145.3 MB.
Access and availability
Weights are on Hugging Face at Cactus-Compute/whistle under Apache 2.0, run with the Cactus needle engine (pip install cactus-needle). Cactus provides prebuilt engines for 17 targets including mobile, browser and WASI.
Official access documentationWhat you can build
Test it on short, representative clips on the target device, measuring word errors, timestamps, latency, and performance on names before using transcripts to trigger actions.
Limitations
Cactus reports Whisper base still leads on TED-LIUM, AMI and the MLS average. Transcripts are capped at 320 tokens, input must be 16 kHz mono, and music can still come out as text, so Cactus recommends a voice activity detector for always-on microphones.