Whistle: Speech Recognition Model for CPU Released on HuggingFace

On September 30, 2026, the model Cactus-Compute/whistle was published on HuggingFace with the automatic-speech-recognition task. The author describes it as a speech-to-text model for mobile devices, wearables, robots, smart homes, cars and microcontrollers.

What it means

The model is distributed under the apache-2.0 license. The largest weights file — checkpoints/whistle.safetensors — takes up 0.2 GB, while the author claims the entire model is packed into a single 16.9 MB file. The weights are published in safetensors format. It runs via the cactus-needle engine, which, according to the author, works on CPU without a GPU and without external dependencies. Whistle performs three tasks on-device: transcription of 16 kHz mono audio up to 30 seconds long in seven languages — Russian is not mentioned among them — English, German, French, Spanish, Italian, Dutch and Polish; output of timestamps for each word with probabilities; extraction of a speech embedding — the encoder output, one row per every 80 ms frame. The model reuses the .cact container, Cactus Quants quantization, SIMD kernels and the KV cache from Needle, so on a device where Needle already runs, Whistle launches in the same binary without a second runtime. The decoder depth can be selected at load time with the --audio-depth flag. For those looking for speech recognition tools that run on your own hardware without a GPU, Whistle is one of the candidates with an open license and claimed CPU support. You can follow the arrival of similar models in the CPU3D news feed.