Рақамли технологияларнинг назарий ва амалий масалалари 9-том 1-нөмір (2026) · 80-89-беттер
Performance evaluation of speaker identification algorithms using speech signal features
Шукуров, К.Э., Хасанов, У.К.
Аңдатпа
This article analyzes the effectiveness of using different models in speaker recognition processes and selects the best one for the system. In terms of accuracy and speed performance of the system, the classical MFCC + cosine similarity and modern x-vector, ECAPA-TDNN + PLDA architectures are compared. Based on the data set generated from different speakers, the accuracy, f1-score, EER, latency, and GPU load indicators of the models are evaluated. According to the experimental results, the ECAPA-TDNN model outperforms the other models with an accuracy of 95.7%. Since the speaker recognition stage is also important for speaker separation systems, accuracy indicators are of high relevance. The ECAPA-TDNN + PLDA model offers good solutions in terms of using computational resources, working with large data sets, and analyzing their data.
идентификация говорящегоECAPA-TDNNx-векторMFCCлогарифмическое сходствоPLDAкосинусное сходствоглубокое обучениевектор признаковречевая биометрия
Метадеректер дереккөзі: журналдың OAI-PMH архиві · Sindex толық мәтінді сақтамайды, дереккөзге сілтеме береді.