Browse Prior Art Database

Discrete Parameter Markov Model Speech Recognition Using Soft Decision Quantization

IP.com Disclosure Number: IPCOM000039153D
Original Publication Date: 1987-Apr-01
Included in the Prior Art Database: 2005-Feb-01
Document File: 2 page(s) / 34K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+3]

Abstract

In a discrete parameter Markov model speech recognition system, a space is partitioned into a plurality of exclusive regions based on specified speech-related characteristics, such as power spectrum features. Each region is represented as a spectral prototype. Input speech is examined at periodic intervals, the speech for each interval representing a frame F. Based on the speech-related characteristics of each frame, an ordered list of the closest M out of the N total prototypes is formed for each frame. Two frames are differentiated if their rank-ordered lists differ; the spectral space will be quantized into N!/(N-M)! regions. Hence, for each transition (i.e., arc) of a Markov model, there must be a probability distribution for N!/(N-M)! possible outcomes.

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 54% of the total text.

Page 1 of 2

Discrete Parameter Markov Model Speech Recognition Using Soft Decision Quantization

In a discrete parameter Markov model speech recognition system, a space is partitioned into a plurality of exclusive regions based on specified speech-related characteristics, such as power spectrum features. Each region is represented as a spectral prototype. Input speech is examined at periodic intervals, the speech for each interval representing a frame F. Based on the speech-related characteristics of each frame, an ordered list of the closest M out of the N total prototypes is formed for each frame. Two frames are differentiated if their rank- ordered lists differ; the spectral space will be quantized into N!/(N-M)! regions. Hence, for each transition (i.e., arc) of a Markov model, there must be a probability distribution for N!/(N-M)! possible outcomes. This results in a large parameter size for determining the value of P(F¯TR), the probability of a frame F given a specific transition TR in a Markov model. The present invention proposes an algorithm for reducing parameter size according to the equation: For a given frame F, an ordered list of M label outputs is represented by (f1, f2,..., fM). A label output-producing transition in a Markov model is referred to as TR. Because the present invention con siders a number of possible label outputs (rather than one) for a frame, the decision quantization is defined as "soft" rather than "hard". The probability of F being produced...