Ragheb, A., Gody, A., Said, T. (2021). Comparative Study of different types of RNN in Speech Classification. The Egyptian Journal of Language Engineering, 8(1), 1-16. doi: 10.21608/ejle.2021.45203.1014
Ayat N. Ragheb; Amr Gody; Tarek Said. "Comparative Study of different types of RNN in Speech Classification". The Egyptian Journal of Language Engineering, 8, 1, 2021, 1-16. doi: 10.21608/ejle.2021.45203.1014
Ragheb, A., Gody, A., Said, T. (2021). 'Comparative Study of different types of RNN in Speech Classification', The Egyptian Journal of Language Engineering, 8(1), pp. 1-16. doi: 10.21608/ejle.2021.45203.1014
Ragheb, A., Gody, A., Said, T. Comparative Study of different types of RNN in Speech Classification. The Egyptian Journal of Language Engineering, 2021; 8(1): 1-16. doi: 10.21608/ejle.2021.45203.1014
Comparative Study of different types of RNN in Speech Classification
1Electronics and communication, Faculty of engineering, Fayoum university, Fayoum, Egypt
2Faculty of Engineering, Fayoum University
3Electronics and communication, Faculty of engineering, Fayoum University, Fayoum, EGYPT
Abstract
This paper introduces different models for pre-processing classification and their performance in Automatic Speech Recognition system. Different Recurrent Neural Network (RNN) architectures have been tested for this problem, such as RNN cells (RNN), bidirectional RNN (BRNN), Long Short-Term Memory (LSTM), and bidirectional LSTM. Mainly two features have been considered. First, Mel frequency cepstral coefficient (MFCC) plus delta and delta-delta coefficients (39 parameters) have been used. Second, MFCC quantization using Vector Quantization technique has been used as features. All models have been trained on TIMIT database. Vowels, nasals, Fricatives, plosives, and silences have been chosen as syllable classes for classification. Experiment results show that BRNN-MFCC-5- {30,30,20,25,25} and BLSTM-MFCC -4- {30,30,25,20} systems with MFCC plus delta and delta-delta coefficients (39 parameters) give the highest accuracy. It achieved 92.6% and 92.07%, respectively. vowels, nasals, and silences give the highest accuracy in BLSTM-MFCC -4- {30,30,25,20}model with 98.5%, 83.6% and 93.7%, respectively. Fricatives and plosives in BRNN-MFCC-5- {30,30,20,25,25} model with 89.7% and 66%, respectively.