Large Vocabulary Arabic Continuous Speech Recognition using Tied States Acoustic Models

Abstract

The Hidden Markov Model (HMM) lies at the heart of the modern speech recognition systems as it provides a simple, effective and straight forward frame work to model the time varying acoustic features of the speech signals. The basic process of building HMM based speech recognition systems is a straight forward process. Nevertheless, the proper parameter estimation of such models requires large training data. Therefore, parameter tying techniques were developed to reduce the parameters of HMMs without affecting the overall system performance. This study proposes an Arabic phonetic decision tree necessary to build Tied State tri-phone HMMs. Experimental results show promising word correctness when compared with both data driven tri-phone models and phoneme based models. The maximum word correctness achieved by the proposed approach was 95.13%. Whereas it reached 78.03 and 58.45% using data driven tri-phones and phoneme based HMMs, respectively, when tested on the same benchmark database.

Arabic speech recognition using end‐to‐end deep learning

“…We compared our best results with previous works by Azim et al [63,64] and Ali et al [27] using WER and word…”

“…We compared our best results with previous works by Azim et al [63,64] and Ali et al [27] using WER and word accuracy (Wacc). Our results achieved WER and Wacc better than their system as in Table 6.…”