Overview
Hush is an open-source noise suppression model developed by weya AI, designed to clean real-time call audio by removing background noise, competing voices, and audio interference. It runs on standard CPUs with minimal latency (0.9 ms per 10 ms frame) and is optimized for Voice AI stacks, improving ASR accuracy and agent experience in noisy environments.
BaseRT is an open-source LLM runtime built specifically for Apple Silicon. It claims to be the fastest runtime on M-series chips, outperforming MLX and llama.cpp in both decode and prefill speeds. With a single-command installation, it allows developers to serve and run local models (Qwen, Llama, Gemma, etc.) entirely on-device, enabling private AI inference for coding agents and other applications.
Feature Comparison
| Feature | Hush | BaseRT |
|---|---|---|
| Primary Function | Real-time noise suppression for voice AI calls | Fast LLM runtime for on-device inference on Apple Silicon |
| Target Use Case | Cleaning noisy call audio for ASR, bots, and agents | Running large language models locally for coding agents and apps |
| Platform | CPU (no GPU needed), deployable in cloud or data center | Apple Silicon (M-series chips) |
| Performance | 0.9 ms per 10 ms audio frame on CPU | Up to 33% faster decode than MLX/llama.cpp; up to 6.4x faster prefill |
| Model Size | ~8 MB | Varies by LLM (e.g., Qwen3 0.6B, Llama 3.2 3B) |
| Open Source | Yes (open-source noise suppression model) | Yes (runtime is open source, models are open) |
| Installation | Pilot program via website; GitHub available | Single curl command |
| Community | Trusted by BFSI teams; Hugging Face leaderboard | Discord community for on-device AI engineers |
Pricing
Hush: The core model is free and open source. weya AI offers a 2-week pilot program for enterprises to integrate Hush into their workflow, with measurement and lift analysis before full rollout.
BaseRT: The runtime is free and open source. There are no API costs; users bring their own models (many are open source) and run them locally on their Apple Silicon devices.
Pros and Cons
Hush
Pros:
- Real-time noise suppression on CPU with low latency
- Small model size (8 MB) easy to deploy
- Open source with strong community backing
- Improves ASR accuracy in noisy environments
Cons:
- Narrow focus on voice call audio cleaning only
- Requires integration into existing Voice AI stack
- Pilot program may not suit all users immediately
BaseRT
Pros:
- Fastest LLM runtime on Apple Silicon
- Single-command installation, easy to use
- Supports popular open models (Qwen, Llama, Gemma)
- Enables fully local, private AI inference
Cons:
- Limited to Apple Silicon hardware
- No built-in noise suppression or audio processing
- Requires separate coding agent setup for full workflow
Verdict
Choose Hush if your primary need is cleaning noisy call audio for voice AI agents or ASR systems, especially in real-time environments. Choose BaseRT if you want to run large language models locally on Apple Silicon for coding agents or other on-device AI tasks. Both are open source but serve very different purposes.