A data reduction method to estimate vowel distributions and its use in comparing two formant estimation methods

Speech features such as formants of vowels uttered by many talkers are considered to form a normal distribution in each phoneme on a feature space. However, those features may apparently show the different dispersions peculiar to the estimation methods. Therefore, if the correct distributions can be found by a credible method, it will make clear the definition of feature estimation errors so that the comparative evaluations between the feature estimation methods will become possible. In this paper, we first propose the data reduction method to estimate true formant distributions of vowels. In the method, we apply the principal component analysis to the formant data of each vowel on a F1-F2 space to search an average value and a three-sigma ellipse. If the average and the ellipse are searched iteratively after removing the outside data of the ellipse regarded as errors, they finally converge. The proportion of the data samples within the final ellipse to all data will be different in the formant estimation methods. We consider that the estimation method of larger proportion is higher in the accuracy because of the high trust. IFC (Inverse Filter Control) method, in which formants are estimated from zero-crossing information, has been compared with LPC method under the above criteria. As a result of the analysis using vowels in words, it has been shown that the IFC method is superior to the LPC in the proportion and the ratio of area in the final ellipse to that in initial one. The proportions of data within the final ellipses are 90-96% in IFC and 84-95% in LPC, which are obtained from five Japanese vowels uttered by 20 males. The intuition obtained by observing the states of distributions supports the numerals in the analysis. Based on the results, we conclude that the formant estimation using zero-crossing information (IFC) is more effective than that by spectral shapes (LPC).