Modern Standard Arabic speech disorders corpus for digital speech processing applications

Digital speech processing applications including automatic speech recognition (ASR), speaker recognition, speech translation, and others, essentially require large volumes of speech data for training and testing purposes. Although there are available speech corpora, speech data for speakers suffering speech disorders are hardly available for many languages including Arabic language. Consequently, developing digital speech processing applications that target the entire society becomes hard due to the unavailability of speech corpora that contain sufficient speakers’ variations including healthy and disordered speech. This research presents our work towards developing a Modern Standard Arabic (MSA) speech corpus for speakers suffering distortion and substitution articulation disorders. The speech corpus was recorded by 40 (20 male and 20 female) Jordanian speakers who suffer either distortion or/and substitution articulation disorders. This speech corpus can be used for various applications including ASR, speech and hearing, and others. Part of this speech corpus is used for developing and evaluating an ASR for MSA using the Carnegie Mellon University (CMU) Pocketsphinx tools based on Mel-Frequency Cepstral Coefficients (MFCC) and Hidden Markov Model (HMM) techniques. Furthermore, Linear Discriminant Analysis (LDA) and Maximum Likelihood Linear Transform (MLLT) optimization techniques were applied. Using three different testing data sets, this work obtained 98.38% and 1.76% average word recognition correctness rate (WRCR) and average Word Error Rate (WER), respectively, for speaker-dependent and text-independent. For speaker-independent and text-dependent, this work obtained 99.37% and 0.68% average WRCR and average WER, respectively, whereas for speaker-independent and text-independent this work obtained 96.53% and 4.00% average WRCR and average WER, respectively.