Improved Statistical Speech Segmentation Using Connectionist Approach

Problem statement: Speech segmentation is an important part for speech recognition, synthesizing and coding. Statistical based approach detects segmentation points via computing spectral distortion of the signal without prior knowledge of the acoustic information proved to be able to give good match, less omission but lot of insertion. These insertion points dropped segmentation accuracy. Approach: This study proposed a fusion method between statistical and connectionist approaches namely the divergence algorithm and Multi Layer Perceptron (MLP) with adaptive learning for segmentation of Malay connected digit with the aim to improve statistical approach via detection of insertion points. The neural network was optimized via trial and error in finding suitable parameters and speech time normalization methods. The best neural network classifier was then fusion with divergence algorithm to make segmentation. Results: The results of the experiments showed that the best neural network classifier used learning rate of value 1.0 and momentum rate of value 0.9 with data normalization based on zero-padded. The segmentation using fusion of statistical and connectionist was able to reduce insertion points up to 10.4% while maintaining match points above 99% and omission point below 0.7% within time tolerance of 0.09 second. Conclusion: The result of segmentation using the proposed fusion method indicated potential use of connectionist approach in improving continuous segmentation by statistical approach.