AP22684173 – Development of a highly efficient neural network method for detecting voice activity at a low signal-to-noise ratio

Purpose: Development of a highly efficient neural network method and training of deep neural networks for detecting voice activity at a low signal-to-noise ratio

Relevance: In the context of the active development of voice technologies and increased requirements for information security, especially at low signal levels, the creation of a highly efficient VAD system based on deep neural networks is becoming an extremely urgent task. This will significantly improve the accuracy of voice activity recognition in a noisy environment and ensure reliable biometric identification.

Scientific supervisor: Ph.D., Aigul Nurlankyzy

Expected and achieved results: 1) Experimental studies necessary to determine the number of training epochs; 2) Conducting experimental studies to select the most appropriate activation function; 3) Experimental studies to conduct a comparative analysis of parameters (training accuracy, validation accuracy, test accuracy) and MLP, CNN, RNN.