Developing a word separation algorithm for recognition of continuous Bangla speech based on prosodic features

Abstract

In the age of information technology Human-Computer interaction has gained importance. Speech is the primary mode of communication among human being and people expect to exchange natural dialect with computer. This expectation can be achieved due to recent development of speech technology. Speech recognition is the process of extracting necessary information from input speech signal to make correct decision and can be applied in automation of operator assisted services, dictation, interactive voice response, medical transcription, pronunciation in computer aided language learning application, data entry etc. To achieve this at first word separation from continuous speech is needed. An isolated speech recognition system requires that a speaker offers clear signature between words but continuous speech consists of continuous utterance which is the representative of a real speech. The thesis develops a word separation algorithm named Prosody based word separation algorithm (PWSA) for recognition of continuous Bangla speech based on prosodic features. Bangla is a bound stress language i.e. it has stress which is high on initial word of a sentence and becomes low at the end of sentences. Based on relative fundamental frequency estimation, PWSA is developed to separate words from continuous Bangla speech. Mel Frequency Cepstral Coefficient (MFCC) is used to extract feature from each separated word. Vector Quantization is used to build codebook of each word. Codebook of all words make database of Bangla speech. For recognition of unknown speech at first PWSA is applied to separate words. MFCC features are extracted from unknown words and compared with database. Experimental result shows that the proposed word separation algorithm using stress information with energy performs excellent. Experiment was performed on 1755 words with 98% accuracy which is 32% better than the existing algorithm . Fo r recognition, the system obtained accuracy of 82%.