New method renders speech in environmental recordings 97.9% unintelligible

A team of researchers from CNRS and Cerema (Modan Tailleur, Mathieu Lagrange, Pierre Aumond, Vincent Tourre) has published a method on arXiv that tackles a growing privacy problem: environmental sound recordings often contain intelligible speech, preventing open sharing of acoustic datasets. Their technique, described in the paper "Enforcing Speech Content Privacy in Environmental Sound Recordings using Segment-wise Waveform Reversal," reverses small waveform segments to distort speech while preserving the integrity of the underlying acoustic scene.

The method uses a pipeline of voice activity detection and speech separation to precisely target speech segments, then applies random waveform reversals. The team achieved a 97.9% Word Error Rate on the processed speech (meaning almost nothing is understandable), while sound source classification accuracy dropped only 2.7%, and the Fréchet Audio Distance (a perceptual quality metric) reached 1.40, indicating high audio quality. An ablation study confirmed each component's contribution. Additionally, random splicing made the method more robust against attempts to recover the original speech, at a slight cost in audio quality. This work could enable researchers to share urban sound recordings (traffic, birds, construction) without violating privacy, opening up valuable datasets for acoustic analysis.