Thesis Details
Evaluation and Optimization of Computational Costs in Speaker Recognition Systems
The goal of this thesis is to propose an evaluation metric that includes computational costs. Computational costs generally do not pose a problem in research, but it can become problematic in a commercial production system, where speed is essential. The proposed metric extends existing evaluation framework from NIST and adds parameter for time unit and time unit cost. These metrics are applied on real ASV and experiments show the potential for further research and possible use. The experiments focus on reducing the computational cost by posing a limit on maximum length of the utterance, but also limiting number of frames for x-vector extraction. Both optimizations reduced the computational costs and reached favorable results for the new metrics. Finally, experiments' results are compared and each system modification is ranked according to the new metrics.
speaker recognition systems, speaker recognition evaluation, speaker recognition optimization, computational costs, evaluation metrics
Češka Milan, doc. RNDr., Ph.D. (DITS FIT BUT), člen
Jaroš Jiří, doc. Ing., Ph.D. (DCSY FIT BUT), člen
Orság Filip, Ing., Ph.D. (DITS FIT BUT), člen
Rychlý Marek, RNDr., Ph.D. (DIFS FIT BUT), člen
@bachelorsthesis{FITBT23008, author = "Sab\'{i}na Gregu\v{s}ov\'{a}", type = "Bachelor's thesis", title = "Evaluation and Optimization of Computational Costs in Speaker Recognition Systems", school = "Brno University of Technology, Faculty of Information Technology", year = 2020, location = "Brno, CZ", language = "english", url = "https://www.fit.vut.cz/study/thesis/23008/" }