Date of Award

9-1992

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Electrical and Computer Engineering

First Advisor

Dennis W. Ruck, PhD

Abstract

In the co-channel speaker separation problem, the goal is to recover two separate speech signals from a monaural channel which contains the sum of the two speech signals. A new methodology is developed that if given that a segment of co-channel speech is separated into a stronger and weaker segment, the correct assignment of these separated segments to the appropriate talker can be made using a Linear Predictive Coding (LPC) based minimum-prediction residual computation. The uniqueness of the developed technique is that no a priori information is required of the co-channel speech signal. The information needed to appropriately assign these separated segments from the co-channel speech signal are clean speech that is separate from the co-channel speech signal that are used to compute model LPC vectors. This clean speech is derived from the same channel that the co-channel speech signal is derived from. This technique has shown the ability to correctly assign the given stronger and weaker segments to the appropriate talker at signal-to-signal ratios down to equal power levels. The resulting separated speech is clearly understandable, and the interfering talker's speech signal is effectively eliminated.

AFIT Designator

AFIT-GE-ENG-92S-07

DTIC Accession Number

ADA256443

Comments

The author's Vita page is omitted.

Download

Share

COinS