Privacidade em diarização do orador: deteção segura de - quem fala quando (PRIVADIA)
Type: National Project
Duration: from 2021 Jan 01 to 2021 Dec 31
Financed by: FCT
Prime Contractor: INESC-ID (Other)
The widespread use of devices with internet access, together with the increasing availability of cloud computing platforms and recent advances in AI, has given rise to a multitude of Machine Learning as a Service applications. With their increasing popularity has also come increasing awareness that they also potentially compromise users' privacy, as shown by the intense public debate that led to, and followed, the creation of GDPR. Among other data types, speech stands out by its potential applications and sensitive nature. Current machine learning models are able to, among many other applications, automatically transcribe speech recordings, identify speakers, and perform "diarization", often referred to as the problem of determining “who spoke when” in a conversation. Privacy in diarization is the focus of the Privadia project. This project addresses this problem by combining state-of-the-art diarization methods (typically based on speaker embedding obtained from the hidden layers of deep neural networks) with cryptographic techniques.