Automatic recognition of liquids and glides
R.L. Kashyap, Mansi Mittal · The Journal of the Acoustical Society of America · 1975
This study describes a new algorithm to recognize liquids (/l/, /r/) and glides (/w/, /y/) spoken by a number of speakers. To make the recognition problem more general, /i/ and /u/, the closest vowels to the glides, are also included in the study. We consider the sequence of time samples of each phoneme as an auto-regressive (AR) process of order r. These r coefficients (estimated by a least-square technique) and signal-to-noise ratio of this process will be taken as raw features for the phoneme under consideration. These raw features contain most of the information (speaker-dependent, speaker-independent, and noise, etc. ) that was contained in the phoneme. Clearly, only the speaker-independent information is required to carry out the above task. To extract the speaker-independent information and reduce the dimension of the raw feature space, we perform a linear transformation. The resulting features (in the lower-dimensional space) are then used for classifying the phonemes by a decision theoretic method, namely, the nearest-mean classification rule. Each of the six phonemes was spoken 10 times by each of the four speakers (two male and two female). We chose r as 8. The correct classification rate of about 85% was observed.