NoizzOff – Real-Time Edge AI Noise Reduction for Robust Voice Processing

NoizzOff is CandyVoice’s real-time AI-powered noise reduction technology, designed to restore intelligible speech in extreme acoustic environments while preserving voice identity, timbre and speech characteristics.

Unlike computationally intensive cloud-based speech enhancement approaches, NoizzOff is designed for fully local Edge AI execution. Its compact 330K-parameter neural network operates on low-power ARM processors with NPUs, enabling real-time processing directly on embedded devices without cloud connectivity.

NoizzOff delivers more than 32 dB of adaptive noise reduction, up to +14.3 dB speech intelligibility improvement, and only 16 ms end-to-end latency, processing audio in 5 ms frames. Its single-microphone architecture facilitates integration into existing devices and communication systems.

The technology can suppress severe environmental noise such as engines, rotors, wind and industrial noise while isolating the voice of interest in complex acoustic environments. By providing a cleaner speech signal, NoizzOff can also improve the quality of input data supplied to downstream speech AI systems such as ASR, voice analytics and other voice-processing applications.

For Horizon Europe collaborative projects, CandyVoice can contribute NoizzOff as a technological building block for embedded and distributed AI, edge computing, robust speech processing and human-machine interaction, particularly where computing resources, connectivity, latency or energy consumption are constrained.

Potential research and integration activities include:

• real-time speech enhancement on constrained edge devices;

• integration of compact AI models into heterogeneous distributed computing architectures;

• robust voice interfaces for AI-enabled systems;

• preprocessing of speech for downstream AI and data-processing applications;

• evaluation across challenging real-world acoustic environments;

• optimisation of AI models for low-power processors and NPUs.

CandyVoice is interested in joining European consortia where NoizzOff can be integrated, validated and further developed within edge AI, distributed computing and next-generation digital systems.