A Study on Isolated-Word Myanmar Speech Recognition via Artificial Neural Networks

Abstract

Speech is an easiest way to communicate with each other. Digital processing of speech signals is very important for speedy and precise automatic speech recognition systems. Speech recognition is the capability of an electronic device to understand spoken words, i.e. the process of decoding an acoustic speech signal captured by a microphone or a mobile phone to a set of words. It is a technology that can be useful in many applications of our daily life, e.g. mobile communications, and has also become a challenge towards human-computer interfacing (HMI) technology. This thesis aims to develop an efficient speech recognition system for isolated Myanmar words based on the theories of digital signal processing, speech processing, and artificial neural network techniques. The proposed system is intended to achieve speaker dependent recognition as well as speaker independent recognition. A speech signal is combined with voice and unvoiced sounds. In addition, each word in the speech is typically surrounded with silence, which may be a hindrance for successful speech recognition. So firstly in this system, the input speeches are manually preprocessed by using the Audacity software in order to detect the start and end points of words and remove unwanted parts like silences in speeches. This system then extracts the acoustically representative features like Mel-Frequency Cepstral Coefficients from the preprocessed speech signals. Finally, those features are used to train a recognition model of neural network with the Backpropagation algorithm for classification and recognition of input speeches. Based on the knowledge learned during training, the recognition model is expected to recognize the same speech by untrained new speakers (i.e. speaker independent recognition). The proposed system in this thesis is developed to recognize twenty isolated Myanmar words, which are the names of the cities in Rakhine state, Shan state, and Kachin state in Myanmar. This system consists of a database which is made up of training and testing data sets with 2400 and 400 utterances respectively. The training words are uttered by 10 speakers (4 males and 6 females) who are university graduate students. As for speaker independent recognition, testing utterances are the same words as in training but uttered by different speakers than the ones participated in training. Theiv proposed system is implemented in MATLAB and experimental results show that it achieved the recognition rate of about 93.5% for known speakers (i.e. speaker dependent) and 76.5% for unknown speakers (i.e. speaker independent).