Optimization and Practical Evaluation of Speaker Verification Models for Resource-Constrained Devices

Nurimov, Parakhat

Al-Farg'oniy avlodlari · 2026-yil

Annotatsiya

This paper investigates the effectiveness of quantization techniques for speaker verification models on resource-constrained Android devices. Tiny-SV-CNN and EdgeSpeakerCNN-AAM were trained on VoxCeleb1, evaluated using the VoxCeleb1-O protocol, and deployed on a Samsung A04 smartphone with TensorFlow Lite. Experimental results show that INT8 significantly reduces the model size of Tiny-SV-CNN, while FP16 preserves the accuracy of EdgeSpeakerCNN-AAM with nearly 50% model size reduction. The study provides practical guidelines for selecting appropriate deployment formats for edge AI applications.

Maqola ma’lumotlari
MualliflarNurimov, Parakhat
JurnalAl-Farg'oniy avlodlari
Nashr sanasi2026-07-19
Son3
Betlar44-48
TilRus

Kalit so‘zlar

Speaker Verification, Android, TensorFlow Lite, kvantlash, FP16, INT8, Edge AI, Tiny-SV-CNN, AAM-Softmax, Speaker Verification,Android,TensorFlow Lite,quantization,FP16,INT8,Edge AI,Tiny-SV-CNN,AAM-Softmax

Ilmiy soha

Al-Farg'oniy avlodlari jurnalidan boshqa maqolalar

Al-Farg'oniy avlodlari — barcha maqolalar