Cactus Compute releases Whistle, a compact 16.9 MB speech recognition model

Cactus Compute has released Whistle, a compact, open-weight speech recognition model designed for edge hardware. At just 16.9 MB, the model is significantly smaller than industry alternatives like OpenAI’s Whisper Base, which is approximately 145.3 MB.

Whistle is designed for local processing on CPUs without external dependencies, making it suitable for smartwatches, wearables, IoT devices, robotics, and web browsers. It supports seven languages: English, German, French, Spanish, Italian, Dutch, and Polish. The model utilizes a C++ engine and the .cact container format, and it is available under the Apache-2.0 license via Hugging Face.

Despite its small footprint, Whistle demonstrates high performance. In benchmarks such as LibriSpeech test-clean, it achieved a 4.31 word error rate, outperforming the larger Whisper Base model's 8.16. The model is optimized for low latency, reaching the first token in 11 milliseconds, and is compatible with 17 target environments including macOS, Linux, Android, iOS, and WebAssembly.