Arabic Speech Recognition Model with BAIDU Deep and Cluster Learning

Fawaz Al-Anzi^*S T Bibin Shalini

• Kuwait University, Kuwait City, Kuwait

This work involves extracting the spectrum from the Arabic raw, unlabeled audio signal and producing Mel-Frequency Cepstral Coefficients (MFCCs). The clustering algorithm groups the retrieved MFCCs with analogous features. The K-means clustering technique was important to our research, performing an unsupervised categorization of the unlabeled Arabic audio data. Employing K-means on the extracted MFCC features allowed us to classify acoustically similar segments into distinct groups without prior knowledge of their characteristics. This initial phase was crucial for understanding the inherent diversity among our diverse sampled dataset. Dynamic Time Warping (DTW) and Euclidean Distance are utilized for illustration. Classification algorithms such as Decision Tree, , XG Boost, KNN, and Random Forest are used to classify the various classes obtained based on clustering. This study also demonstrates the efficacy of Mozilla's deep speech framework for Arabic speech recognition. The core component of deep speech is its neural network architecture, which consists of multiple layers of Recurrent Neural Networks(RNNs). It strives to comprehend the intricate patterns and interactions between spoken sounds and their corresponding textual representations. The clustered labeled Arabic audio dataset along with transcripts and Arabic Alphabets, is used as input to the Baidu deep speech model for training and testing purposes. PyCharm in Python 3.6 is used to build a Dockerfile. Creating, editing, and managing Dockerfiles within the IDE is made simpler by PyCharm's functionality and integrated environment. Deep speech provides an eminent Arabic speech recognition quality with reduced loss, word error rate, and character error rate. Baidu's deep speech intends to achieve high performance in both end-to-end and isolated speech recognition with good precision and a low word rate and character error rate in a reasonable amount of time. The suggested strategy yielded a loss of 276.147, a word error rate of 0.3720, and a character error rate of 0.0568. This technique increases the accuracy of Arabic automatic speech recognition(ASR).