Adapting parameter-efficient fine-tuning methods to self supervised learning models for Turkish speaker verification

This thesis addresses the problem of adapting large-scale self-supervised learning (SSL) models, which are predominantly pre-trained on English data, to the task of Turkish speaker verification with high performance and low computational cost. The primary objective of this study is to systematically investigate the effectiveness of parameter-efficient fine-tuning (PEFT) methods on state-of-the-art SSL models, namely WavLM and Wav2Vec2. In this context, various PEFT techniques such as adapters, low-rank adaptation (LoRA), and prompt tuning, along with their combinations, were experimented on using the Common Voice Turkish dataset, and the results were benchmarked against a strong ECAPA-TDNN baseline model. For evaluation, a novel composite score metric was utilized, balancing the equal error rate (EER) and the minimum detection cost function (min-DCF). The experimental results demonstrate that the WavLM model adapted with the prompt tuning method achieved the best performance on Common Voice Turkish test set, with an EER of 13.63% and a combined score of 1.303. This result outperformed the best ECAPA-TDNN scenario (13.92% EER and 1.454 combined score), proving that SSL models and PEFT methods constitute a superior approach for overcoming the challenges of language mismatch and data scarcity in low-resource languages like Turkish. This study reveals that PEFT is not only a computationally efficient alternative but can also yield better performance than full fine-tuning by preventing the catastrophic forgetting phenomenon. However, on a YouTube-based test set designed to measure the models' robustness to different data distributions, SSL models achieved higher error rates compared to ECAPA-TDNN. This finding demonstrates the difference in generalization performance between the models, the sensitivity of performance to the characteristics of the test data, and the importance of dataset scale. Consequently, this thesis sets a new state-of-the-art performance for Turkish speaker verification and provides the first comprehensive study that systematically benchmarks PEFT methods in this domain.