Speech recognition method and apparatus

• US 5,091,947 A
• Filed: 08/13/1990
• Issued: 02/25/1992
• Est. Priority Date: 06/04/1987
• Status: Expired due to Term

• Current Assignee: Ricoh Company Limited
• Original Assignee: Ricoh Company Limited
• Inventors: Kawamoto, Toshiki, Kuriki, Syoji, Ariyoshi, Takashi, Yasuda, Seigou, Nakatani, Tomofumi
• Primary Examiner(s): Shaw, Dale M.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US07/566,884
• Time in Patent Office: 561 Days
• Field of Search: 364/531.5, 381/41-45
• US Class Current: 704/246
• CPC Class Codes: G10L 15/32 Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

A voice recognition apparatus includes a coefficient memory for storing at least one coefficient for correcting a degree of similarity which is obtained by either one of the speaker-independent recognition and the speaker-dependent recognition. The apparatus also includes a voice identification circuit for comparing the degrees of similarity of the candidates obtained by either one of the speaker-independent recognition and the speaker-dependent recognition with corrected degrees of similarity of the candidates related to the other recognition type which are obtained by performing a predetermined operation to the degree of similarity of each candidate which is supplied from the other recognition type. Then the voice identification decides one candidate having the highest degree of similarity to be an identification result.

Claims

1. A speech recognition apparatus comprising:

feature extracting means for extracting features of an unknown input voice and producing a corresponding unknown voice pattern;

speaker-independent reference pattern storing means for storing speaker-independent reference patterns;

speaker-dependent reference pattern storing means for storing speaker-dependent reference patterns;

speaker-independent recognition means for calculating a degree of similarity between the unknown input pattern and each of the speaker-independent reference patterns and producing at least one candidate for the unknown input pattern;

speaker-dependent recognition means for calculating a degree of similarity between the unknown input pattern and each of the speaker-dependent reference patterns and producing at least one candidate for the unknown input patterns;

coefficient storing means for storing at least one coefficient for correcting the degree of similarity which is obtained by either one of the speaker-independent recognition means and the speaker-dependent recognition means;

correcting means for multiplying said coefficient and the degree of similarity of said candidate supplied from a predetermined one of said speaker-independent recognition means and said speaker-dependent recognition means to generate corrected degrees of similarity of candidate; and

identification means for comparing the degree of similarity of said candidate supplied from the predetermined one of the speaker-independent recognition means and the speaker-dependent recognition means with said corrected degree of similarity of said candidate related to the other recognition means supplied from said correcting means and for selecting one candidate having the highest degree of similarity among said degree of similarity of candidate and said corrected degree of similarity of candidate, said selected one candidate being the recognized result.

2. The speech recognition apparatus as claimed in claim 1, further comprising confirmation means for outputting the candidate having the highest degree of similarity to a speaker who utters the voice in order to confirm whether or not the candidate is correct.

3. The speech recognition apparatus as claimed in claim 2, wherein the identification means further comprises cancelling means for cancelling the candidate when it is confirmed to be incorrect;

candidate selecting means for selecting other candidates one by one from among the candidates in decreasing order of degree of similarity; and

coefficient adjusting means for adjusting a value of the coefficient when one candidate which is selected by the candidate selecting means and is confirmed to be correct by the confirmation means is one of the reference patterns to which the cancelled candidate does not belong, so that one of the reference patterns to which the cancelled candidate does not belong is selected as the first candidate when the same input voice is processed.

4. The speech recognition apparatus as claimed in claim 1, wherein the value of the coefficient is decided on the basis of the experiment.

5. The speech recognition apparatus as claimed in claim 1, further comprising adding means for adding a predetermined value to said corrected degree of similarity of candidate.

6. The speech recognition apparatus as claimed in claim 1, wherein the coefficient has a plurality of different values and wherein the speech recognition apparatus further comprises means of manually setting one of the predetermined values of the coefficient which is to be multiplied.

7. The speech recognition apparatus as claimed in claim 2, wherein the speaker-independent reference pattern storing means for storing a reference voice with respect to a public organization, and wherein the speech recognition apparatus further comprises telephone number storing means for storing telephone numbers;

and communication means for communicating with a communication apparatus, and controlling means for selecting one telephone number corresponding to the recognition result which is output by the identification means from among the telephone numbers stored in the telephone number storing means, so that the selected telephone number is sent to the communication apparatus.

8. A speech recognition method comprising the steps of:

producing an unknown input voice pattern of an unknown input voice from features extracted from the unknown input voice;

calculating a degree of similarity between the produced unknown pattern and reference patterns which are composed of reference patterns used exclusively for speaker-independent recognition and reference patterns used exclusively for speaker-dependent recognition to thereby generate at least one candidate related to said speaker-independent recognition and at least one candidate related to said speaker-independent recognition;

correcting the degree of similarity of said candidate a predetermined one of the speaker-independent recognition and the speaker-dependent recognition by multiplying said degree of similarity and a coefficient to thereby generate corrected degree of similarity of candidate;

comparing said degree of similarity of said candidate related to the predetermined one of said speaker-dependent recognition and said speaker-dependent recognition with said corrected degree of similarity of said candidate related to the other recognition; and

selecting one candidate having the highest degree of similarity among said degree of similarity of candidate and said corrected degree of similarity of candidate, said selected one candidate being the recognized result.

9. The speech recognition method as claimed in claim 8, further comprising a step of outputting the candidate having the highest degree of similarity to a speaker who utters the voice in order to confirm whether or not the candidate is correct.

10. The speech recognition method as claimed in claim 9, further comprising steps of cancelling the candidate when it is confirmed to be incorrect;

selecting other candidates one by one from among the candidates in decreasing order of degree of similarity; and

adjusting a value of the coefficient when one candidate which is selected by the selecting step and is confirmed to be correct by the confirmation step is one of the reference patterns to which the cancelled candidate does not belong, so that one of the reference patterns to which the cancelled candidate belongs is not likely to be selected at the following recognition stage.