Evaluation Of Unsupervised Models Using Minimal Pair ABX Measure

Y, Satya Dheeraj (2015) Evaluation Of Unsupervised Models Using Minimal Pair ABX Measure. Masters thesis, Indian Institute of Technology Hyderabad.

Preview

Text
EE13M1009.pdf - Submitted Version
Download (1MB) | Preview

Abstract

The Minimal-Pair ABX (MP-ABX) task has been proposed as a method for evaluating speech features for zero resource (i.e only limited amount of labelled data) unsupervised speech technologies. MP-ABX task is an alternative to the phoneme word error rate,it is necessary to discriminate between the minimal pair of words from a language. We compared Mel Frequency Cepstral Coefficients (MFCC) with modelling parameters of these MFCC's by using unsupervised generative models like Gaussian Mixture Model (GMM) and Gaussian-Bernoulli Restricted Boltzmann Machine (GBRBM). In an MP-ABX task, the features (MFCC) a, b and x associated to three speech sounds, A, B and X are computed, where A and B are chosen to be minimally different words (e.g. dog vs doll) and X is linguistically identical to either A or B, although it can be indexically different (different talker or added noise). Then, one determines whether x is closer to a or b by computing Distance Time Wrapping algorithm (DTW) of the evaluated features. By repeating this on a representative set of A, B, X triplets, a measure of the discriminability of minimal pairs when coded with the tested featural representation is obtained. This evaluation metric is especially suitable for zero-resource settings

[error in script]

IITH Creators: