Speaker: Almudena Aguilera. Abstract: In this study we will evaluate the discriminatory behaviours that are generated in speaker recognition systems, specifically those that verify whether two audios belong to the same speaker or not. These systems work by extracting the unique characteristics of each individual, allowing their identification.
With the raised use of the speaker recognition systems, concerns had increased about the possible discrimination that these systems generate in different biased groups based on protected characteristics, such as the gender, age, accents, etc. From this recent concern about the possible lack of fairness between groups, arises the main motivation of this work.
During the development of this work, we will analyze the possible biases that are generated in speaker recognition systems based on DNN-Embeddings, for the following sensitive characteristics: gender, accent and age. To carry out the analysis, the two databases commonly used in these kind of systems, Voxceleb and Mozilla Common Voice, will be employed; in the case of the second database, the FairVoice subset will be used.
In the development of this work, we have analyzed the discriminatory behaviours of some already trained models. Additionally, we have generated several models using a data preprocessing technique, for assess if these techniques would reduce the difference of precision that take place between different groups.