Browse Prior Art Database

AN EFFICIENT DATA STORAGE AND RETRIEVAL SYSTEM FOR DISCRETE HIDDEN MARKOV MODEL SPEECH RECOGNITION SYSTEMS

IP.com Disclosure Number: IPCOM000007110D
Original Publication Date: 1994-Feb-01
Included in the Prior Art Database: 2002-Feb-26
Document File: 5 page(s) / 246K

Publishing Venue

Motorola

Related People

Ed Srenger: AUTHOR [+4]

Abstract

We propose here a novel approach for storing and retrieving the statistical data associated with a HMM by taking advantage of existing redundancy and clustering properties of that data. This method results in a significant reduction in the required mem- ory storage requirements while leaving the overall recognition accuracy of the system unalfected adding only minimally to the overall computational load.

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 27% of the total text.

Page 1 of 5

MOliOROLA Technical Developments Volume 21 February 1994

AN EFFICIENT DATA STORAGE AND RETRIEVAL SYSTEM FOR DISCRETE HIDDEN MARKOV MODEL SPEECH RECOGNlTlON SYSTEMS

by Ed Srenger, Bill Kushner, Matt Hartman and Orhan Karaali

  We propose here a novel approach for storing and retrieving the statistical data associated with a HMM by taking advantage of existing redundancy and clustering properties of that data. This method results in a significant reduction in the required mem- ory storage requirements while leaving the overall recognition accuracy of the system unalfected adding only minimally to the overall computational load.

resolution in a given state (e.g. the number of state events chosen to represent a set ofacoustic features), and, of course, the total size of the vocabulary to be recognized. Reducing the first two parameters usu- ally produces a degradation in recognition accuracy, while a reduction ofthe vocabulary size might limit the scope ofthe application under consideration.

I. SUMMARY OF THE INVENTION

III. DESCRIPTION OF THE INVENTION

II. BACKGROUND OF THE INVENTION

  Current speech recognition systems fall into two broad classes: deterministic and statistical. Both approaches require that memory storage be provided for reference information used in the pattern matching procedure yielding the recognized utterance.

  In a statistical based system each vocabulary item is characterized by a Hidden Markov model (HMM) whose topology is described by a set of states and allowed transitions between states (cf Figure 1). To each state is assigned a set ofprobabilities represent- ing the likelihood of different acoustic events occur- ring in that state. In addition, probabilities are assigned to the likelihood oftransiting between states of the model. In the method described herein, it will be assumed that each HMM will fully repre- sent an entire vocabulary word. This is to distin- guish from systems that might be phoneme based and require a set of phone HMMs to Lilly charac- terize a vocabulary item.

  The amount of memory required to store the statistical information associated with a HMM, the state event probabilities and state transition probabilities, can rapidly become prohibitive for a real-time hardware implementation. The amount of information that needs to be stored can be controlled to some extent by varying several parameters: the number of states in the HMM, the acoustic event

0 Moforola, 1°C. 1994

  In a standared discrete HMM structure, the amount data required to represent the event probabilities for each state is significantly larger than the relatively smaller number of probabilities that are needed to characterize the possible transitions between states.

  Let us define the number of states in a discrete HMM model by N and the number of events (or symbols) per state as M (where usually M >> N). For a discrete HMM system M corresponds to the size of a codebook table. In a discrete HMM sys- tem, the incoming acoustic fea...