Speech recognition

• US 20040117182A1
• Filed: 09/25/2003
• Published: 06/17/2004
• Est. Priority Date: 04/19/2001
• Status: Active Grant

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Downey, Simon N
• Granted Patent: US 7,970,610 B2
• Time in Patent Office: Days
• US Class Current: 704/243
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning
• CPC Class Codes: G10L 2015/228 of application context
• CPC Class Codes: H04M 2201/40 using speech recognition
• CPC Class Codes: H04M 3/4931 Directory assistance systems
• CPC Class Codes: H04M 3/4936 Speech interaction details speech recognition per se G10L15/00

Abstract

In this invention vocabulary size of a speech recognizer for a large task is reduced by providing a recognizer only for the most common vocabulary items. Uncommon items are catered for by providing aliases from the common items. This allows accuracy to remain high while also allowing uncommon items to be recognized when necessary.

Claims

1. A method of speech recognition comprising the steps of:

a) comparing a first audio signal comprising a first unknown utterance with a first set of audio representations to generate a first measure of similarity for each audio representation of said set,

each audio representation being associated with a corresponding first item of data, a first item of data being associated with an associated item of data, the associated item of data having an audio representation which is not one of said set;

b) comparing a second audio signal comprising a second unknown utterance with a second set of audio representations to generate a second measure of similarity for each audio representation of said second set, each audio representation of said second set being associated with a corresponding second item of data

c) selecting from data defining associations between items of data, items of data which are defined as being associated with one another according to

the first item of data for which the first generated measure indicates the greatest similarity;

an item of data associated with the first item of data; and

the second item of data for which the second measure indicates the greatest similarity.

2. A method according to claim 1 in which the associated item of data is determined according to the steps of

d) comparing a known utterance to said first set of audio representations to generate a third measure of similarity for each of said representations;

and

e) associating an item of data which characterises the known utterance with the audio representation for which the generated third measure indicates the greatest similarity.

3. A method according to claim 1 in which the associated item of data is generated according to the steps of

f) comparing a sequence of reference models representing an item of data to a plurality of sequences of reference models representing the audio representations of the first set in order to generate a measure of similarity for each of said plurality of sequences;

and

g) associating the item of data with the audio presentation represented by the sequence of reference models for which the generated measure indicates the greatest similarity.

9. A device as claimed in any preceding claim wherein the database stores a plurality of records each of which includes the name of a customer as an item of data of the first category.

10. A carrier medium carrying processor implementable instructions for causing a processor to carry out the steps of any of claims 1 to 5 during implementation of the instructions.

5. A method according to claim 4 in which the comparing step f) uses a confusion matrix which characterises errors which occur in the comparing step a) for said audio representations.

6. A device for retrieving a data record from a database storing a plurality of data records each of which includes a data item of a first category and a data item of a second or subsequent category, wherein the data items in the first category are designated as being either common or uncommon in dependence upon the frequency with which they appear in the data records stored in the database, the device comprising:

audio representation storage means for storing an audio representation in respect of each of the common data items in the first category;

association storage means for storing associations between each common data item and a plurality of uncommon data items whose audio representations are similar to but different from the audio representation of the respective associated common data item;

comparison means for comparing a signal derived from an unknown utterance with each of the audio representations of common data items stored in the audio representation storage means, generating a measure of similarity at least in respect of one or more audio representations which are sufficiently similar to the compared signal to give rise to a measure of similarity above a predetermined threshold and designating as candidate first category data items both the common data items whose audio representations gave rise to a measure of similarity above the threshold and the uncommon data items associated with the designated common data items according to the association storage means;

selection means for selecting one or more data items of a second or subsequent category; and

retrieval means for retrieving one or more data records including a first category data item equal to one of the candidate first data items designated by the comparison means and a second or subsequent category data item selected by the selection means.

7. A device according to claim 6 wherein the comparison means includes a speech recognition device connected to a public switched telephone network for receiving the signal via the public switched telephone network from a user using a terminal connected to the network, said user uttering the unknown utterance.

8. A device according to claim 6 wherein the selection means also includes a speech recognition device connected to a public switched telephone network for receiving the signal via the public switched telephone network from a user using a terminal connected to the network, said user uttering the unknown utterance.