Method of training a speaker-dependent speech recognizer with automated supervision of training sufficiency

Claim 2

A process for training a speaker dependent speech recognizer to recognize a new voice message comprising the steps of, (a) converting a new voice message spoken by a user into a token, (b) comparing said token to a plurality of templates stored in said recognizer to obtain a recognition score for each of said templates, said templates comprising templates of voice messages said recognizer has been previously trained to recognize, and one or more previously formed templates of said new voice message, (c) applying first, second and third tests to said recognition scores to determine if said token and one of said previously formed templates of said new voice message can become templates of a voice message now recognized by said recognizer, if said new voice message is too close to one of said voice messages said recognizer has been previously trained to recognize, or if said user should repeat said new voice message to form another token, wherein said first test comprises determining if the best of said recognition scores is greater than a first predetermined threshold, said second test comprises determining if the difference between the best of said recognition scores with respect to templates of voice messages said recognizer has been previously trained to recognize and the best of said recognition scores with respect to said one or more previously formed templates of said new voice message is greater than a second predetermined threshold, and said third test comprises determining if the best of said recognition scores with respect to said one or more previously formed templates of the new voice message is smaller than a third predetermined threshold.

Claim 3

The process of claim 1 wherein at least first and second successive tokens of said new voice message are formed, wherein said first token is compared to the templates of the voice messages said recognizer has been previously trained to recognize to obtain recognition scores, wherein the first test but not the second and third tests is applied to the recognition scores of the first token, and wherein when the second token of said new voice message is formed, said first token is a previously formed template of said new voice message.

Claim 4

A process for training a speaker dependent speech recognizer comprising the steps of (a) converting a new voice message spoken by a user into a token, (b) comparing said token to a plurality of templates stored in said recognizer to obtain a recognition score for each of said templates, said templates comprising templates of voice messages said recognizer has been previously trained to recognize and one or more previously formed templates of said new voice message, (c) testing said recognition scores to determine if said new voice message is too similar to a voice message said recognizer has been previously trained to recognize, if said user is not using sufficiently consistent pronunciation in successive pronunciations of said new voice message, or if said speech recognizer can now recognize said new voice message.

Claim 5

A process for training a speaker dependent speech recognizer comprising the steps of (a) converting a new voice message spoken by a user into a token, (b) comparing said token to a plurality of templates stored in said recognizer to obtain a recognition score for each of said templates, said templates comprising templates of voice messages said recognizer has been previously trained to recognize and one or more previously formed templates of said new voice message, (c) testing said recognition scores to determine if said new voice message is too similar to a voice message said recognizer has been previously trained to recognize, if said user is not using sufficiently consistent pronunciation in successive pronunciations of said new voice message, or if said speech recognizer can now recognize said new voice message.

Claim 6

The method of claim 3 wherein said testing step comprises determining if the difference between the best of said recognition scores with respect to said templates of said voice messages said recognizer has been previously trained to recognize and the best of said recognition scores with respect to said one or more previously formed templates of said new voice message is greater than a predetermined threshold, and determining if the best of said recognition scores with respect to the one or more previously formed templates of the new voice message is smaller than another predetermined threshold.

Claim 7

The method of claim 4 wherein said testing step further comprises determining if the best recognition score with respect to all of said templates is greater than a predetermined threshold.

Claim 8

A process for training a speaker dependent speech recognizer comprising the steps of: (1) prompting a user to speak a new voice message that the user currently desires the recognizer to be trained to recognize for the i th time, i≧2, (2) converting said new voice message spoken for the i th time into a new token, (3) comparing the new token to templates of voice messages that the recognizer has been previously trained to recognize and one or more previously formed templates of the new voice message to obtain a recognition score of the token with respect to all of said templates, (4) storing the new token as a template, (5) applying a first test which is passed if the best of said recognition scores is greater than a first threshold (6) determining if the best recognition score was obtained with a template of a voice message the recognizer has been previously trained to recognize or with a previously formed template of the new voice message, (7) if the best recognition score was obtained with a previously formed template of the new voice message, (a) applying a second test which is passed if the difference between the best recognition score made with a template of a voice message the recognizer was previously trained to recognize and best recognition score made with a previously formed template of said new voice message exceeds a second threshold, (b) performing a third test which is passed if the best recognition score falls below a third threshold, (c) if said second and third tests are passed and if the template which produced the best recognition score passed the first test at the time it was the token, then said new token and said template with said best recognition score are stored as first and second templates of the new voice message, which new voice message can now be recognized by said recognizer (d) if said second test is not passed and said third test is passed, or if said second and third tests are passed but said template with the best recognition score has not passed the first test at the time it was a token, prompting the user with a message that the new voice message is too similar to a voice message that the recognizer has been previously trained to recognize.

Claim 9

A process for training a speaker dependent speech recognizer comprising the steps of: (1) prompting a user to speak a new voice message that the user currently desires the recognizer to be trained to recognize for the i th time, i≧2, (2) converting said new voice message spoken for the i th time into a new token, (3) comparing the new token to templates of voice messages that the recognizer has been previously trained to recognize and one or more previously formed templates of the new voice message to obtain a recognition score of the token with respect to all of said templates, (4) storing the new token as a template, (5) applying a first test which is passed if the best of said recognition scores is greater than a first threshold (6) determining if the best recognition score was obtained with a template of a voice message the recognizer has been previously trained to recognize or with a previously formed template of the new voice message, (7) if the best recognition score was obtained with a previously formed template of the new voice message, (a) applying a second test which is passed if the difference between the best recognition score made with a template of a voice message the recognizer was previously trained to recognize and best recognition score made with a previously formed template of said new voice message exceeds a second threshold, (b) performing a third test which is passed if the best recognition score falls below a third threshold, (c) if said second and third tests are passed and if the template which produced the best recognition score passed the first test at the time it was the token, then said new token and said template with said best recognition score are stored as first and second templates of the new voice message, which new voice message can now be recognized by said recognizer (d) if said second test is not passed and said third test is passed, or if said second and third tests are passed but said template with the best recognition score has not passed the first test at the time it was a token, prompting the user with a message that the new voice message is too similar to a voice message that the recognizer has been previously trained to recognize.

Claim 10

The process of claim 6, further comprising the step of: if the best recognition score was obtained with a template of a voice message that the recognizer has been previously trained to recognize and if said first test was not passed and if said third test was passed by said new token, prompting the user with a message that the new voice message is too similar to a voice message that the recognizer has been previously trained to recognize.

Claim 11

The process of claim 6, further comprising the step of: if the best recognition score was obtained with a previously formed template of a new voice message, and the second test is not passed and the third test is not passed, or the second test is passed and the third test is not passed, returning to step (1) so long as i is below a maximum limit.

Claim 12

The process of claim 6, further comprising the step of: if the best recognition score was obtained with a template of a voice message that the recognizer has been previously trained to recognize, and if said first test was passed, returning to step (1) so long as i is below a maximum limit.

Claim 13

The process of claim 6, wherein for the case i=1, the process comprises the steps of: prompting the user to say the new voice message, converting said new voice message into a token, comparing the token to stored templates of voice messages that the recognizer has been previously trained to recognize to obtain a recognition score with respect to all of these templates, storing the token as a template, applying said first test which is passed if the best of said recognition scores is greater than said first threshold, and prompting the user to say the current voice message again.