Method of learning in a speech recognition system

• US 6,138,097 A
• Filed: 09/28/1998
• Issued: 10/24/2000
• Est. Priority Date: 09/29/1997
• Status: Expired due to Term

• Current Assignee: RPX Clearinghouse LLC (RPX Corporation)
• Original Assignee: Nortel Networks Corporation
• Inventors: Lelievre, Laurent, Glorion, Catherine, Lockwood, Philip
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Wieland, Susan
• Application Number:
  US09/161,589
• Time in Patent Office: 757 Days
• Field of Search: 704/256, 704/258, 764/255
• US Class Current: 704/256.2
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/0638 Interactive procedures

Abstract

A recognition test matches a speech segment supplied to the system with a set of parameters associated with a reference and memorized in a dictionary. A provisional version of each set of parameters to be memorized in the dictionary in association with a reference is estimated on the basis of one or more segments of speech, after which the provisional version is memorized in the dictionary in association with the reference. At least one repetition of the speech segment is submitted to a recognition test, after which depending on whether it has matched the speech segment with the provisional version, the provisional version is modified and the modified provisional version is memorized.

Claims

1. In a speech recognition system comprising a dictionary for storing parameter sets associated with respective references and means for carrying out recognition tests, wherein each recognition test matches a speech segment supplied to the system with at least one parameter set from said dictionary, a learning method for generating an entry in said dictionary in association with a reference, which comprises obtaining a plurality of speech segments each pronounced by a speaker to correspond to said reference, and processing said plurality of speech segments, said processing comprising the steps of:

/a/ estimating a provisional version of a parameter set on the basis of at least one of said plurality of speech segments;

/b/ storing said provisional version in the dictionary in association with said reference;

/c/ submitting at least one of said plurality of speech segments to a recognition test; and

/d/ modifying the provisional version stored in a manner depending on whether said recognition test matches said speech segment with said provisional version or not.

2. A method according to claim 1, comprising the steps of:

obtaining m speech segments each pronounced by the speaker to correspond to said reference, wherein m is an integer at least equal to 2;

estimating said provisional version of the parameter set on the basis of said m speech segments;

carrying out steps /c/ and /d/ for at least one of said m speech segments.

3. A method according to claim 1, wherein the parameter sets stored in said dictionary define hidden Markov models.

4. A method according to claim 1, wherein after said parameter set has been estimated from the plurality of speech segments pronounced by the speaker, a sound restitution of one of said plurality of speech segments is performed.

5. A method according to claim 1, comprising the steps of:

obtaining n speech segments each pronounced by the speaker to correspond to said reference, wherein n is an integer at least equal to 1;

estimating said provisional version of the parameter set on the basis of said n speech segments;

prompting the speaker to pronounce a supplementary speech segment corresponding to said reference;

submitting said supplementary speech segment to a recognition test;

updating the provisional version stored in the dictionary if the recognition test matches said supplementary segment with the provisional version of the parameter set.

6. A method according to claim 5, further comprising repeating at least once the sequence of said steps of prompting, submitting and updating.

7. A method according to claim 5, further comprising the step of submitting each of said n speech segments to a recognition test.

8. A method according to claim 6, further comprising the step of delivering a message to the speaker to indicate a risk of confusion, when at least two recognition tests to which different speech segments are submitted match said different speech segments with a same parameter set from the dictionary other than said provisional version.

9. A method according to claim 6, wherein the provisional version of the parameter set is definitively recorded in the dictionary, in association with said reference, after it has been updated a determined number of times.

10. A method according to claim 6, further comprising the step of replacing the provisional version stored in the dictionary by a new provisional version of the parameter set, estimated on the basis of at least one of two consecutive speech segments, when both recognition tests to which said consecutive speech segments are submitted fail to match said consecutive speech segments with the provisional version of the parameter set.