Privacy preserving and personalized transformations for speech recognition
PhD Student: Pierre Champion
Advisor(s): Denis Jouvet (Loria), Anthony Larcher (LIUM - LST)
Funding: ANR Deep-Privacy
The objective of the proposed thesis is twofold. First, it aims at finding a privacy preserving transform of the speech data, and, second, it will investigate the use of additional personalized transforms, that can be applied on the user’s terminal, to increase speech recognition performance.
In the proposed approach, the device of each user will not share its raw speech data, but a privacy preserving transformation of the user speech data. In such approach, some private computations will be handled locally, while some cross-user computations may be carried out on a server using the transformed speech data, which protect the speaker identity and some of his/her features (gender, sentiment, emotions…). More specifically, this will rely on a representation learning to separate the features of the user data that can expose private information from generic ones useful for the task of interest, i.e., here, the recognition of the linguistic content. We will build upon ideas of Generative Adversarial Networks (GANs) for proposing such a privacy preserving transform. Since a few years, GANs are getting more and more used in deep learning. They typically rely on both a generative network and a discriminative network, where the generator aims to output samples that the discriminator cannot distinguish from the true samples [Goodfellow et al., 2014; Creswell et al., 2018]. They have also been used as autoencoders [Makhzani et al., 2015] which are made of three mains blocks: encoder, generator and discriminator. In our case, the discriminators shall focus on discriminating between speakers and/or between voice-related classes (defined according to gender, emotions, etc). The training objective will be to maximize the speech recognition performance (using the privacy preserving transformed signal) while minimizing the available speaker or voice-related information measured by the discriminator.
As devices are getting more and more personal, it creates opportunities to make speech recognition more personalized. This includes two aspects: adapting the model parameters to the speaker (and to the device) and introducing personalized transforms to help hiding the speaker voice identity. Both aspects will be investigated. Voice conversion approaches provide example of transforms aiming at modifying the voice of a speaker so that it sounds like the voice of another target speaker [e.g., Chen et al., 2014; Mohammadi & Kain, 2014]. Similar approaches can thus be applied to map speaker specific features to those of a standard (or average) speaker, which thus would help concealing its identity. To take benefit of the increased personal usage of terminals, speaker and environment specific adaptation will be investigated to improve speech recognition performance. Collaborative learning mixing speech and speaker recognition has been shown to benefit both tasks [Liu et al. 2018; Garimella et al. 2015] and provide a way to combine both information in a single framework. This approach will be compared to adaptation of deep neural networks-based models [e.g., Abdel-Hamid & Jiang, 2013] to handle best different amounts of adaptation data.