Hush

Hush is a pragmatic choice for developers building voice agents in noisy environments. Its real-time CPU performance is genuinely impressive—sub-millisecond per frame means no latency spike. The open-source model is freely available on GitHub and Hugging Face, so you can self-host or embed it however you like. Where it shines: call centers with background chatter, field ops with street noise, any scenario where ASR accuracy dips because of audio quality. We'd reach for Hush when we need a lightweight preprocessor that works without a GPU. But it has limits. Hush is not a plug-and-play consumer app; you'll need to integrate the model into your pipeline. It focuses on speech noise suppression, not general audio cleaning (music, sound effects). For teams wanting a fully managed service, Weya AI's platform includes noise cancellation as a built-in feature. Compared to proprietary noise suppression (like NVIDIA's RTX Voice), Hush is free and runs on CPU, but may not match GPU-backed quality in extreme noise. The model is also tied to Weya AI's ecosystem for direct integration, though the open-source release mitigates lock-in. In practice, the 8MB model size makes it easy to deploy even on edge devices. The lack of recent news means it's stable but not rapidly evolving. For most voice AI teams, Hush is a smart, cost-free low-hanging fruit to improve call quality.