Feature Fusion Extraction Method for Improvement of Recognition of Continuous Speech : A Feature Fusion Method for Recognition of Continuous Speech

To interact with a computer speech recognition plays a vital role. Speech recognition is a difficult task due to phonological variation, individual differences, environmental factors such as background noise, diversity of languages used, and real-world issues such as accents. Speech recognition of audio mathematical expression is also challenging due to symbols, Greek letters, digits, and operators present in it. Feature extraction plays a very important role in speech recognition. In speech recognition, MFCC is the most commonly used feature extraction method but it has some weaknesses. It works well for small datasets but for large datasets, accuracy goes down. This paper presents a feature fusion based methodology for the recognition of speech for mathematical expression. Feature level fusion is performed where time, frequency, and cepstral domain features are combined which improves the result of speech recognition. The dataset consists of 6 types of mathematical expressions recorded by 10 people. A total of 600 .wav files are given as input to train the model. The proposed feature extraction model is trained using a sequential classifier and KNN classifier. The result indicates the proposed feature extraction algorithm gives better accuracy than using only the time domain, frequency domain, and cepstral domain for both sequential classifier and KNN.