Development of Digital Signal Processing and Statistical Classification Methods for Distinguishing Nasal Consonants

Abstract

For almost half a century, people have been looking for efficient classifiers to distinguish two nasal sounds, / / from / /, uttered by a single speaker. From the middle of the last decade, there has been little progress in research on this topic. In recent years, we, researchers of the Voice I/O Group in Department of Computer Science at North Carolina State University, have conducted some new trials on this classical problem. In this thesis, those trials are briefly summarized. Instead of simply using the Fourier transform to produce the spectra as people usually did in the past, the author uses other kinds of transforms to extract more feature differences between / / and / /. The new transforms can be the alternatives of frequencies, such as singular values or eigenvalues, or even other transforms such as wavelets, which can deal with non-stationary systems quite well. We combine together the old and new features to get a larger feature vector, which will bring more classification information. We collect multiple voice samples of a single speaker and calculate the above feature representations, then use them as input of some popular statistical classification techniques, such as Principle Component Analysis (PCA), Discriminant Analysis (DA), and Support Vector Machine (SVM). By way of one training process, one testing process, and one heuristic scheme, we can identify the nasals with low error rates.