Question

Compare MFCC with PLP.

Answer 1

Introduction

In traditional phonology, the place and the manner of the articulation in the vocal tract are applied to

classify Chinese initials. And the statistical and psychological methods are used to explore the

perceptual characteristics. The characteristics of phonation and articulation such as voiced or

voiceless, aspirated or unaspirated, and fricative or frictionless, are the most important factors that

influences the perception of initials [1, 2]. A perceptual measurement based on LPC among Chinese

finals has been proposed in [3], which makes it easier to evaluate the equivalence of different

audiometric word lists. The acoustic features most commonly used are Mel Frequency Cepstral

Coefficients (MFCC) and Perceptual Linear Prediction (PLP) features [4]. Both MFCC and PLP are

tested with and without ‘pitch’ information using the same back-end on an English consonants corpus

and the results are compared with human listener results at the level of articulatory feature

classification, which shows that no representation reaches the levels of human performance but PLP

has higher accuracies for most manner values on English consonants than MFCC [5]. However, the

perception of Chinese initials, which are not exactly the same as English consonants, is more difficult

for humans, especially for patients, than that of Chinese finals. Hence, it is very important to do the

research on the perceptual characteristics of Chinese initials.

In this paper, we discuss which acoustic features or their combinations are the most consistent with

the perception of Chinese initials. We systematically test PLP and MFCC representations of Chinese

initials by carrying out two experiments with respect to acoustic space and perceptual space,

respectively. We then combine the results of the two experiments by using a statistical method, called

Spearman's rank correlation coefficient, to assess how well the relationship between two types ofdistance can be described using a monotonic function. We also single out a proper acoustic feature

representations for Chinese initials and distance metrics between different categories of initials to

measure the acoustic distance which is monotonically related to the perceptual distance.