Browse Prior Art Database

Subsampled Prototypes and Weighted Clustering

IP.com Disclosure Number: IPCOM000038506D
Original Publication Date: 1987-Jan-01
Included in the Prior Art Database: 2005-Jan-31
Document File: 3 page(s) / 31K

Publishing Venue

IBM

Related People

Cohen, JR: AUTHOR [+2]

Abstract

The present invention relates to the characterization of the support of the distribution of speech by prototypical points obtained through vector quantization biased to represent a uniform distribution on the support. One method of characterizing speech is by prototypical points (in Euclidean space) determined from a clustering algorithm. The present invention proposes two algorithms which seek to cover the support of speech (like a phonetician) without speech-theoretic notions or biasses. The algorithms are constructed to automatically produce a set of prototypes whose distribution represents, not the complex bumpy distribution of real speech, but the uniform distribution on the support of real speech.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 3

Subsampled Prototypes and Weighted Clustering

The present invention relates to the characterization of the support of the distribution of speech by prototypical points obtained through vector quantization biased to represent a uniform distribution on the support. One method of characterizing speech is by prototypical points (in Euclidean space) determined from a clustering algorithm. The present invention proposes two algorithms which seek to cover the support of speech (like a phonetician) without speech-theoretic notions or biasses. The algorithms are constructed to automatically produce a set of prototypes whose distribution represents, not the complex bumpy distribution of real speech, but the uniform distribution on the support of real speech.

The reason such algorithms are required is that the support of speech inherits the complexity of the speech distribution and therefore cannot be described in a comprehensive, geometric or algebraic way. Moreover, this complexity also prohibits a simple Monte Carlo procedure for the uniform distribution. The algorithms that follow do not require any description of the support. One algorithm selects a subsample of the usual sample so that the subsample is, for example, uniform. The other algorithm uses the usual sample from the distribution of speech but it adjusts the contribution of points in centroid computations so as to achieve, on average, the same effect. The resulting prototypes may be used directly, or they may serve as initial values for a usual k-means iteration on time contiguous speech data. The first algorithm (subsampling) includes the following steps: 1. Draw a simple random sample x1,x2,...,xN from the distribution of speech and use some local density estimate (e.g., based on the k nearest neighbors) for the probability density p(xi), for i = 1,...,N. 2. Select a subsample having the desired distribution. 3. Choose k points at random in the subsample. 4.

With these as initial values, cluster the subsample by the k-means algorithm. For a random variable X with probability (density) function p(x), we define a subsampling indicator function I = I(X) of the form I(X) = 1 if X is to be included in the subsample, I(X) = 0, otherwise. Suppose we wish the subsample to have the properties of a sample from some other p...