Speaker independent Sinhala speech recognition for voice dialling

Speech is the most natural and the most powerful way of communication between humans. Speech recognition for voice dialling applications has already been developed for languages such as English, French and Japanese, etc. However, there is no evidence of existence of such applications for voice dialling in Sinhala language with speaker independent environment. The work described in this paper is based on an attempt to implement a speaker independent Sinhala speech recognizer and a voice dialling application which has been used to communicate with a VoIP application. The underlying concept of building the speech recognizer is Hidden Markov Model (HMM) and the system is developed using the Hidden Markov Model Toolkit (HTK). The first stages of building the speech recognizer involve the preparation of speech samples for training and the creation of the pronunciation dictionary which lists all the speech samples along with their phonetic representations. A noise reduction method has been applied at the front end of the voice dialling application to clean up the speech signal from the beginning. The middle stages comprise of employing a good feature extraction technique to enhance the speech recognition, and building and training the acoustic model to match a spoken digit to the observed input while the latter stages involve the creation of the language model to determine which digit has spoken. The results show that 87.37% of the digits are correctly recognized by the speech recognizer under quiet environment while 82.19% of the digits are correctly recognized in noisy environment.