Automatic phoneme segmentation of Tamil utterances

Speech recognition systems can be designed using sub-word unit phoneme where as phoneme is the smallest natural linguistic unit represents unique sound in particular language. Speech recognition process carried out in two phases: segmentation and recognition. Speech segmentation is an important phase in continuous speech recognition, since it reduces the search space. In larger vocabulary tasks, automatic segmentation of speech utterances into phonemes is preferable than manual segmentation which is a tedious and time consuming one. There are many automatic phoneme segmentation methods like Spectral Transition Measure (STM), Maximum Likelihood Segmentation, Level Building Segmentation, Variable Length Segment Quantization. In the proposed work, automatic segmentation of Tamil speech into phonemes has been carried out using STM and Level Building Dynamic Programming (LBDP). Both the algorithms use spectral variation as the base to find the boundaries of phoneme. A speech corpus of 100 Tamil speech utterances consisting of 25 unique Tamil words is used. Each word is uttered by 4 native speakers of Tamil language. This experiment is carried out after extracting 12 Mel Frequency Cepstral Coefficient (MFCC) of the speech. The performance of the segmentation techniques are measured corresponding to manual segmentation.