Skip to content
DSML.KZ
Publications

Article · Bulletin of the National Engineering Academy of the Republic of Kazakhstan · September 30, 2024

Application of the conformer model for kazakh speech recognition

Orken J. Mamyrbayev, Turdybek Kurmetkan, Recep Sinan Arslan

Orken Mamyrbayev
Open via DOI

Abstract

The article describes the application of a transformer-based conformer model and a convolutional neural network (CNN) in Kazakh speech recognition, with an overview of automatic speech recognition (ASR) technologies. By exploring ways to combine convolutional neural networks and transformers, the best aspects of both models were combined to simulate audio circuits. The structure of the Conformer model is described. To determine the performance of the Conformer model, it was compared with previous Transformer and CNN models. According to the results of the study, the Conformer model outperformed the Transformer and CNN models in recognizing Kazakh speech, and we achieved higher accuracy than our previous studies. During testing, the word error rate (WER) dropped to 12.3 and the letter error rate (SER) dropped to 6.9 percent.

Speech Recognition and SynthesisSpeech and Audio ProcessingMusic and Audio Processingtransformer modelsconvolutional neural networksCNN-Transformer hybrid

More by Orken Mamyrbayev