Sevdi, Onur ErenTopaloglu, YakupKayaarma, Selma Yilmazyildiz2026-02-082026-02-0820259798331597276https://doi.org/10.1109/ASYU67174.2025.11208275https://hdl.handle.net/20.500.12885/52952025 Innovations in Intelligent Systems and Applications Conference, ASYU 2025 -- 2025-09-10 through 2025-09-12 -- Bursa -- 214381Speaker identification plays a key role in various applications, such as security, biometrics, and human-computer interaction. As a specific task under the domain of audio classification, speaker identification aims to recognize individuals based on their voice characteristics. This paper presents a comparison between three widely adopted neural network architectures and evaluates their performance as classifiers for a real-time speaker identification system. A custom-collected dataset was gathered using publicly shared YouTube videos of a single speaker imitating multiple characters from traditional Turkish shadow play Karagoz and Hacivat. Both MFCC and Log-Mel filterbank energy features were used during the training of CRNN, 2D-CNN and Bi-LSTM architectures. Among these architectures, 2D-CNN achieved the highest accuracy with a value of 94.4% and was approximately 2.7 times faster than its closest follower Bi-LSTM during real-time testing on RTX 4070 Super GPU. © 2025 IEEE.eninfo:eu-repo/semantics/closedAccessAudio ClassificationBidirectional Long Short Term MemoryConvolutional Neural NetworksConvolutional Recurrent Neural NetworksSpeaker IdentificationConference Object10.1109/ASYU67174.2025.112082752-s2.0-105022444330N/A