Browse Prior Art Database

Run-Length Adjustment of Hidden Markov Model Parameters for Speech Recognition

IP.com Disclosure Number: IPCOM000039113D
Original Publication Date: 1987-Apr-01
Included in the Prior Art Database: 2005-Feb-01
Document File: 2 page(s) / 13K

Publishing Venue

IBM

Related People

Bahl, LR: AUTHOR [+4]

Abstract

In a Markov model speech recognition system, the probability of each output symbol (or label or feneme) y which can be generated at a model transition (or arc) is adjusted based on the number of times the symbol occurs in a row (a) in actual data and (b) in synthetic data. In speech recognition, speech may be characterized as a spectral space partitioned into a finite number of regions based on prescribed speech features. Each region is identified by a symbol (or a label or feneme). Incoming speech is examined at successive time intervals based on the prescribed speech features. The symbol which is "closest" (by some measure) to the speech features of an interval is assigned to the interval. The incoming speech can thereby be represented as a sequence of symbols, L1 ...

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 2

Run-Length Adjustment of Hidden Markov Model Parameters for Speech Recognition

In a Markov model speech recognition system, the probability of each output symbol (or label or feneme) y which can be generated at a model transition (or arc) is adjusted based on the number of times the symbol occurs in a row (a) in actual data and (b) in synthetic data. In speech recognition, speech may be characterized as a spectral space partitioned into a finite number of regions based on prescribed speech features. Each region is identified by a symbol (or a label or feneme). Incoming speech is examined at successive time intervals based on the prescribed speech features. The symbol which is "closest" (by some measure) to the speech features of an interval is assigned to the interval. The incoming speech can thereby be represented as a sequence of symbols, L1 ...Lm, where each L is selected from an alphabet of symbols y1, y2,..., yn . This sequence, typically generated by an acoustic processor, corresponds to actual data. One approach to speech recognition involves the use of Hidden Markov Models (HMMs). A Markov model includes a plurality of states; transitions, each of which extends from a state to a state and each of which has a probability assigned thereto; and for each symbol which can be generated at a given transition there is a respective output probability. If there are 200 symbols in the alphabet, each transition will normally have 200 output probabilities assigned thereto. The probabilities are determined in a training phase, as discussed in [*]. With trained models, known Monte Carlo techniques can be used to synthetically generate symbol data that corresponds to some given speech input. Such data...