Browse Prior Art Database

Nonlinear Score Adjustment for Speech Recognition

IP.com Disclosure Number: IPCOM000042093D
Original Publication Date: 1984-Mar-01
Included in the Prior Art Database: 2005-Feb-03
Document File: 2 page(s) / 18K

Publishing Venue

IBM

Related People

Cohen, JR: AUTHOR [+2]

Abstract

Nonlinearly combining the information from the language model and information from the acoustic front end increases score differences and improves computer speech recognition. Many recognition errors occur when the language model suggests the correct word, but the acoustic match is very small and incorrect. Since these errors are separable from the correctly decoded words simply on the basis of the acoustic match score, scaling may be accomplished as a function of this match score.

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 100% of the total text.

Page 1 of 2

Nonlinear Score Adjustment for Speech Recognition

Nonlinearly combining the information from the language model and information from the acoustic front end increases score differences and improves computer speech recognition. Many recognition errors occur when the language model suggests the correct word, but the acoustic match is very small and incorrect. Since these errors are separable from the correctly decoded words simply on the basis of the acoustic match score, scaling may be accomplished as a function of this match score. For greater differentiation in the detailed match score, for each word, recalculate the acoustic match log probability p, using the following equations, to provide an adjusted acoustic match log probability p': These equations are derived by mapping the score p to a hyperbola of the form (p-x) (p- kx) = g, after having offset p = 0 to an appropriate position. The process has three parameters, g, k and offset, as follows: g: affects the radius of curvature of the transition from the asymptote of x - y = 0 to the asymptote of x - ky = 0. k: the inverse of the slope of the asymptote for P < 0. offset: the position of the "break" in the hyperbola. Typical values are: g = 5 k = 10 offset = 2

1

Page 2 of 2

2

[This page contains 1 picture or other non-text object]