VoxShield – Real-Time Voice Anonymisation for Privacy-Preserving AI & Digital Communications

VoxShield is CandyVoice’s real-time voice anonymisation technology, designed to protect speaker identity by irreversibly removing identifiable biometric voice characteristics while preserving speech intelligibility and linguistic information.

Voice is a unique biometric identifier and can be used to recognise, track and correlate speakers across datasets and communications. VoxShield addresses this privacy challenge by separating linguistic content from biometric voice identity and generating a non-traceable voice independent of the original speaker.

Unlike approaches relying on speaker embeddings or predefined identity models, VoxShield operates directly on the speech signal and does not require a reference identity model. The technology achieves a drastic reduction in speaker identification capability, with a measured residual similarity of approximately 10⁻⁸.

VoxShield operates in real time with approximately 70 ms latency and is designed for fully embedded Edge execution without cloud dependency. It can run on ARM, x86 and DSP architectures and can also be integrated into server, telecom and secure digital infrastructures.

Importantly for AI applications, anonymised speech remains usable by both humans and automatic speech recognition (ASR) systems without significant degradation. Voice normalisation can also reduce ASR system complexity. This enables voice data to retain much of its operational and linguistic value while substantially reducing exposure of biometric speaker identity.

For Horizon Europe collaborative projects, CandyVoice can contribute VoxShield as a technological building block for privacy-preserving AI, biometric data protection, trustworthy data processing and secure exploitation of voice datasets, including distributed and edge environments.

Potential research and integration activities include:

• privacy-preserving processing and sharing of voice data;

• anonymisation of speech datasets for AI development and training;

• protection of biometric identity in distributed AI systems;

• integration of anonymisation at the edge before data transmission or processing;

• evaluation of privacy, intelligibility and downstream AI performance;

• development of trustworthy voice-processing pipelines compatible with European data-protection requirements.

CandyVoice is interested in joining European consortia where VoxShield can be integrated, validated and further developed as a privacy-enhancing technology for trustworthy AI, data spaces and distributed digital infrastructures.