Whistle: An Ultra-Lightweight Speech-to-Text Tool in 16.9 MB
Cactus Compute released Whistle, an open-source speech-to-text model designed for local execution that requires only a single 16.9 MB file. The model is built to run efficiently on standard CPUs for edge and on-device environments. This extreme compactness enables advanced offline voice processing and speech recognition on resource-constrained hardware such as mobile devices, wearables, and microcontrollers. It significantly lowers the barrier for deploying local, privacy-preserving voice agents without relying on heavy cloud infrastructure. Whistle is published under the Apache-2.0 license and is designed to integrate with the Needle runtime for small on-device language models. Audio input and tool processing remain completely contained within the local engine.
## BACKGROUND
Speech-to-text models typically require substantial memory and computational power, often demanding gigabytes of storage or cloud-based APIs to function accurately. Running these models locally on CPUs or microcontrollers usually involves heavy quantization and optimization to bypass the need for dedicated GPUs.