Researchers at Sweden's Chalmers University of Technology and RISE Research Institutes have proposed a technique that can remove attributes like gender from speech data to protect sensitive information.
The researchers used a generative adversarial network known as PCMelGAN, which features a generator that creates samples and a discriminator that aims to differentiate between the generated samples and real-world samples.
PCMelGAN maps speech recordings to mel spectrograms, representations of the spectrum of frequencies of the audio signal as it varies over time.
It then passes the speech recordings through a filter that removes sensitive information and a generator that adds synthetic information, and inverts the mel spectrogram output into audio in the form of a raw waveform.
The researchers found PCMelGAN "can successfully obfuscate sensitive attributes in speech data and generates realistic speech independent of the sensitive input attribute" while maintaining "a high level of utility."
View Full Article
Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA
No entries found