Browse Prior Art Database

Accuracy improvement of selection of source and destination points in speech recognition

IP.com Disclosure Number: IPCOM000012254D
Original Publication Date: 2003-Apr-23
Included in the Prior Art Database: 2003-Apr-23
Document File: 5 page(s) / 19K

Publishing Venue

IBM

Abstract

Disclosed is a system to improve speech recognition accuracy for the system that needs to ask a user to input source and destination points by speech recognition. The speech recognition is a useful method in the system that needs to ask a user to choose one from many like the system that provides information of the routes between two train stations, departure and arrival time and so forth. Because a user needs to go through many steps to reach source and destination station name that the user wants to input in GUI system. However, many choices sometimes cause bad recognition accuracy, consequently this may degrade usability seriously. Our system improves recognition accuracy by utilizing additional information, that is time required from source to destination point, whether the destination is a major train station or not, the departure time's probability - e.g. departure time at 2AM or 3AM is in low probability - and so on. We can derive these kinds of information from the words defined in the grammar and the recognition results. Our system utilizes these kinds of information to improve recognition accuracy by adding "weight" to the grammar and making corrections on the order of recognition results sorted by the 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 30% of the total text.

Page 1 of 5

THIS COPY WAS MADE FROM AN INTERNAL IBM DOCUMENT AND NOT FROM THE PUBLISHED BOOK

JP820020959 Koichiro Kato/Japan/IBM Yoshinori Tahara, Hiroya Ogihara, Masahiro B Baba

  Accuracy improvement of selection of source and destination points in speech recognition

1. Introduction

Our system improves speech recognition accuracy for the system that needs to ask a user to input source and destination points by speech recognition. The speech recognition is a useful method in the system that needs to ask a user to choose one from many like the system that provides information of the routes between two train stations, departure and arrival time and so forth. Because a user needs to go through many steps to reach source and destination station name that the user wants to input in GUI system. However, many choices sometimes cause bad recognition accuracy, consequently this may degrade usability seriously.

Our system utilizes the additional information that derives from the words defined in the grammar and the recognition results to improve recognition accuracy. We exemplify our system by the system that queries and provides information of the routes between two train stations, departure and arrival time according to user's inputs by speech recognition.

In this system, a user usually speaks "from station A to station B at T o'clock for arrival" for input. The speech recognition engine(herein after SR engine) can return more than one recognition results with a score for the source train station name "A", the destination train station name "B" and the arrival time "T".

We can consider the following two methods to correct the recognized results:

(1) Using weight in a grammar
(2) Making corrections on the order of recognition results sorted by the score as a post process

A grammar is a file that defines valid words or phrases for recognition. We can specify "weight" in the grammar. The method (1) is suitable for the information that is available when a grammar is generated. The method (2) is suitable for using the information that can be derived from the recognized results.

Our system is using both (1) and (2) methods. The system configuration is shown in Fig 1.

Recognition results

Train route query system

Corrected results

Speech Recognition Engine

   Gramar with "weight"

Generate

Module to generate grammar adding "weight"

  Module for Post Process

Words used for grammars(name of station) Data for weight
Data for parameters

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

Page 2 of 5

Example how to weight(Assign value from 0 thru 1 like probability)

¦ ---------

¦ ----

¦ ----

¦ ----

¦ ----

0+----------------------------

1 2 3 4 5 6 .......... The number of connections to other lines

(2) Distance from the current user's position (For source station) ANI(Automatic Number Identification) can be utilized to know user's position in case of a system for telephone.

Example how to weight(Assign value from 0 thru 1 like probability)

The nearest station...