Speaker Verification without Feature Extraction

Speaker verification is a field that is still improving its state of the art (SotA) and tries to meet the demands of its use in speaker authentication systems, forensic applications, etc. The improvements are made by the advancements in deep learning, the creation of new training and testing datasets and various speaker recognition challenges and speech workshops. In this thesis, we will explore models for speaker verification without feature extraction. Inputting the models with raw speaker waveform simplifies the pipeline of the systems, thus saving computational and memory resources and reducing the number of hyperparameters needed for creating the features from waveforms that affect the results. Currently, the models without feature extraction do not achieve the performance of the models with feature extraction. By applying various techniques to the models we will try to improve the baseline performance of the current models without feature extraction. The experiments with SotA techniques improved the performance of a model without feature extraction considerably however we still did not achieve the performance of a SotA model with feature extraction. However, the improvement is considerable enough so that we can use the improved model in a fusion with feature extraction model. We also discussed the experimental results and proposed improvements that aim to solve discovered limitations.

speaker verification, featureless extraction, speaker embedding, residual network, RawNet, VoxCeleb1, VoxCeleb2, VoxSRC, Feature Map Scaling, SincNet, Additive Angular Margin loss, fusion