Discriminative Techniques for Hindi Speech Recognition System

For the last two decades, research in the field of automatic speech recognition (ASR) has been intensively carried out worldwide, motivated by the advances in signal processing techniques, pattern recognition algorithms, computational resources and storage capability. Most state-of–the–art speech recognition systems are based on the principles of statistical pattern recognition. In such systems, the speech signal is captured and preprocessed at front-end for feature extraction and evaluated at back-end using continuous density hidden Markov model (CDHMM). Maximum likelihood estimation (MLE) and several discriminative training methods have been used to train the ASR, based on European languages such as English. This paper reviews the existing discriminative techniques like maximum mutual information estimation (MMIE), minimum classification error (MCE), and minimum phone error (MPE), and presents a comparative study in the context of Hindi language ASR. The system is speaker independent and works with medium size vocabulary in typical field conditions.