Giving a voice to voice privacy
If voice were a colour, it’d be the new black. “From our mobile phones to changing the television channel, we are increasingly using our voices to interact with everyday devices,” says Emmanuel Vincent, senior research scientist and head of science at Inria, France’s National Institute for Research in Digital Science and Technology. However, before Alexa can answer your question, she first needs to be trained on a vast amount of speech and text data. To do this, technology companies typically collect voice data from users and hire human annotators to transcribe them into text. Application developers then define a list of possible user requests and associated answers for every application. The process is then repeated for every language. According to Vincent, this process raises a whole host of problems. “First and foremost, it raises critical privacy and confidentiality concerns relating to the users’ voice characteristics and the spoken content,” he explains. “Furthermore, the process is expensive and, as a result, is inherently not inclusive as many languages and dialects are left out.” With the support of the EU-funded COMPRISE project, Inria is leading an effort to define a new approach to training voice interaction technologies. “By implementing a new methodology and developing new software tools, we created solutions that not only protect privacy and confidentiality, but also reduce costs and increase the inclusiveness of this rapidly growing technology,” adds Vincent, who serves as the project’s coordinator.
Addressing the privacy problem
When the project first started, voice anonymisation was a relatively new concept, meaning there wasn’t much research on it. “Most approaches to assessing anonymisation were based on an assumption that the attacker was naïve and attempted to reidentify the speaker using biometric identification software designed for unprocessed speech,” remarks Vincent. “We quickly realised that this assumption was flawed.” So, to start, researchers rewrote the story. “We defined a stronger attack model where the attacker is aware of the anonymisation system and adapts the biometric identification software accordingly,” says Vincent. From here, the project developed innovative software tools that have reduced the risk of speaker reidentification by several orders of magnitude. Beyond increasing privacy, the project also created solutions that increase inclusiveness by allowing for the development of dialogue systems without the need for training resources in the target language. Furthermore, the project developed a method that has been shown to reduce the cost of integrating voice features into mobile applications by more than 70 %. All these tools are now available to voice technology companies and application developers as open-source via the COMPRISE website.
Bringing voice privacy into the GDPR
In addition to the cutting-edge research and technological advancements achieved by COMPRISE, the project also helped raise awareness about the growing issue of voice privacy. In fact, Vincent has been invited to speak to the French data protection agency and contributed to the European Data Protection Board’s consultation on voice assistants. “I sincerely hope that the GDPR application guidelines will evolve so as to better protect privacy and confidentiality and allow European companies to compete with tech giants,” concludes Vincent. One of those companies is Nijta. The new start-up is a spin-off from the COMPRISE project and aims to offer the secure voice anonymisation solutions businesses need to build trustful relationships with their customers.
Keywords
COMPRISE, voice privacy, voice interaction technologies, voice data, voice anonymisation, data protection, GDPR
