Privacy-Preserving Speech Representation Learning using Vector Quantization

Abstract: Speech signals are a rich source of speaker-related information, including

sensitive attributes like gender, identity, etc. Those sensitive attributes can
be extracted and used for malicious purposes like voice spoofing.
Despite the inherent sensitivity of speech signals, more and more services,
mainly virtual assistants like Apple’s Siri, Google Now, Microsoft’s Cortana,
or Amazon’s Alexa, process, collect and store personal speech signal
in centralized servers raising severe privacy concerns.

The main focus of my work is the investigation of anonymization techniques
to remove sensitive attributes from speech signals while preserving the

linguistic content. In this seminar, I will present my latest approaches that

anonymize the hidden bottleneck representation of an ASR system so that

it cannot be used to identify speakers. The method is based on Vector Quantization,

whose primary goal is to generate a compressed discrete representation

of the input data. I will present how Vector Quantization can learn private

representation, show the results, and discuss future directions.