Cactus Compute releases Whistle, a 16.9 MB on-device speech-to-text model

Cactus Compute released Whistle, an open 16.9 MB CPU speech-to-text model with word timestamps for seven languages and clips up to 30 seconds.

Cactus Compute has released Whistle, an open-weights speech-to-text model that ships as a single 16.9 MB file and runs on the CPU with no GPU. If you're building private dictation, voice notes or a voice front end for a device, this is a transcription model small enough to bundle with the app. Audio stays on the device.

Per the Hugging Face model card, Whistle takes 16 kHz mono audio of up to 30 seconds per pass. It supports English, German, French, Spanish, Italian, Dutch and Polish. It returns word-level timestamps with probabilities, speech embeddings and keyword biasing for names and product terms. Cactus says silence returns an empty transcript rather than an invented sentence. Install is pip install cactus-needle, and the model runs in the same C++ engine as Cactus's Needle tool-calling model. That means a single call can take a clip in and return tool calls.

The benchmarks need some care. Cactus says Whistle beats Whisper base on LibriSpeech and the FLEURS average. It also leads on SPGISpeech and Earnings-22, but Whisper never published figures for those. Whisper base still wins on TED-LIUM, AMI and the MLS average. The Whisper and Moonshine figures come from their authors' papers, not from re-runs. RuntimeWire notes that all results are company-reported. The headline latency, 11.1 ms to first token versus 73.2 ms for Whisper base, was measured only on an Apple M4 Pro. No phone, Raspberry Pi or microcontroller numbers have been published.

It's also unclear how Whistle handles noisy audio or recordings longer than 30 seconds. ArtRealmAI suggests planning your own chunking. The Voxtype integration plan cites Cactus documentation saying telemetry is on by default in the distributed runtime. Verify that before treating Whistle as fully offline. Cactus ships a needle whistle compare command that runs Whistle, Whisper and Moonshine side by side, so you can test it on your own audio and hardware.

Sources