Syllable based word identification for malayalam speech using machine learning
Maya Moneykumar (93514008)
Syllable based word identification for malayalam speech using machine learning
This thesis aims at discussing the development of an isolated word identification system
for Malayalam using Machine Learning techniques. This work examines how Artificial
Neural Network (ANN) and Hidden Markov Model (HMM) can benefit a medium size vocabulary, speaker independent isolated word level recognition system. The goal of this work is to design an ANN based word recognition system and evaluate its accuracy in different modes namely words within the vocabulary as well as out of vocabulary. The
recognition accuracy is also tested for speaker dependent as well speaker independent
modes. The system was then compared with a similar system performance based on
HMM. The work aims at syllable based word identification where each and every utterance will be segmented into corresponding syllables which are in turn trained by the system. Currently, most speech recognition systems are based on Hidden Markov Model (HMM) which is a statistical framework that supports both acoustic and temporal modeling. In this work, the system is trained using syllables segmented from the utterances where a new approach is made to do the syllable segmentation efficiently, based on the energy measure, formant frequencies and zero crossing rate. These segmented syllables are then trained using HMM and ANN to compare the recognition accuracy. To compare the two systems, we have kept similar, the train and test data and also the extracted features. The comparison includes the overall system performance and different test accuracy rates for both the models. The system is trained using multiple utterances of 80 different words by 9 different speakers, 6 male and 3 female, where the feature extraction was done using MFCC, the most powerful feature extraction technique. In this work, the speech recognition engine is built using HTK and WEKA. The work also attempts to evaluate the improvement in recognition accuracy of the system based on ANN by training and testing with additional parameters. The system proved successful in identifying the utterances of out of vocabulary words, which indeed is a notable step in the area of speech recognition.
COMPUTING METHODOLOGIES
ARTIFICIAL INTELLIGENCE
NATURAL LANGUAGE PROCESSING
SPEECH RECOGNITION
Syllable based word identification for malayalam speech using machine learning
This thesis aims at discussing the development of an isolated word identification system
for Malayalam using Machine Learning techniques. This work examines how Artificial
Neural Network (ANN) and Hidden Markov Model (HMM) can benefit a medium size vocabulary, speaker independent isolated word level recognition system. The goal of this work is to design an ANN based word recognition system and evaluate its accuracy in different modes namely words within the vocabulary as well as out of vocabulary. The
recognition accuracy is also tested for speaker dependent as well speaker independent
modes. The system was then compared with a similar system performance based on
HMM. The work aims at syllable based word identification where each and every utterance will be segmented into corresponding syllables which are in turn trained by the system. Currently, most speech recognition systems are based on Hidden Markov Model (HMM) which is a statistical framework that supports both acoustic and temporal modeling. In this work, the system is trained using syllables segmented from the utterances where a new approach is made to do the syllable segmentation efficiently, based on the energy measure, formant frequencies and zero crossing rate. These segmented syllables are then trained using HMM and ANN to compare the recognition accuracy. To compare the two systems, we have kept similar, the train and test data and also the extracted features. The comparison includes the overall system performance and different test accuracy rates for both the models. The system is trained using multiple utterances of 80 different words by 9 different speakers, 6 male and 3 female, where the feature extraction was done using MFCC, the most powerful feature extraction technique. In this work, the speech recognition engine is built using HTK and WEKA. The work also attempts to evaluate the improvement in recognition accuracy of the system based on ANN by training and testing with additional parameters. The system proved successful in identifying the utterances of out of vocabulary words, which indeed is a notable step in the area of speech recognition.
COMPUTING METHODOLOGIES
ARTIFICIAL INTELLIGENCE
NATURAL LANGUAGE PROCESSING
SPEECH RECOGNITION