Recognition confidence measuring by lexical distance between candidates

• US 8,990,086 B2
• Filed: 07/31/2006
• Issued: 03/24/2015
• Est. Priority Date: 02/09/2006
• Status: Expired due to Fees

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Jeong, Sang-Bae, Kim, Nam Hoon, Han, Ick Sang, Choi, In Jeong, Jang, Gil Jin, Jeong, Jae-Hoon
• Primary Examiner(s): Serrou, Abdelali
• Application Number:
  US11/495,562
• Publication Number:
  US 20070185713A1
• Time in Patent Office: 3,158 Days
• Field of Search: 704/251, 704/233, 704/231, 704/256, 704/256.1, 704/252, 704/200, 704/234, 704238-240, 704/270, 704/270.1, 704/275
• US Class Current: 704/251
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams

Abstract

A recognition confidence measurement method, medium and system which can more accurately determine whether an input speech signal is an in-vocabulary, by extracting an optimum number of candidates that match a phone string extracted from the input speech signal and estimating a lexical distance between the extracted candidates is provided. A recognition confidence measurement method includes: extracting a phoneme string from a feature vector of an input speech signal; extracting candidates by matching the extracted phoneme string and phoneme strings of vocabularies registered in a predetermined dictionary and; estimating a lexical distance between the extracted candidates; and determining whether the input speech signal is an in-vocabulary, based on the lexical distance.

Claims

1. A recognition confidence measurement method comprising:

detecting a feature vector of an input speech signal through normalization of a histogram and extracting a phoneme string from the detected feature vector of the input speech signal;

obtaining information about a distance between a phoneme string of the input speech signal and a phoneme string of a predetermined vocabulary, by a predetermined phoneme confusion matrix, wherein the phoneme confusion matrix sets the distance to be decreased in proportion to the increase of the matching degree between phonemes;

extracting as candidates, the phoneme string of the vocabulary that has a higher similarity which denotes a comparatively shorter distance from the speech signal, from a predetermined dictionary;

estimating a lexical distance between the extracted candidates, the estimating of the lexical distance comprising selecting a pair of candidates from the extracted candidates and performing a dynamic matching of the selected pair of candidates; and

determining whether the input speech signal is an in-vocabulary, based on the lexical distance, wherein the dynamic matching comprises;

for each phoneme in a first candidate of the selected pair of candidates, determining as a first matching pair the phoneme in the first candidate and the corresponding phoneme in a second candidate of the selected pair of candidates and determining as a second matching pair the phoneme in the first candidate and a phoneme having an identical shape to the phoneme in the first candidate;

wherein the estimating of the lexical distance further comprises;

calculating a score for the pair of candidates; and

estimating the lexical distance using the calculated score;

wherein the calculating of the score calculates the score using the phoneme confusion matrix.

2. The method of claim 1, wherein the extracting of the phoneme string extracts an optimum phoneme string according to each language by using a Hidden Markov Model (HMM) and a predetermined phoneme grammar for each language.

3. The method of claim 1, wherein estimating the phoneme confusion matrix comprises:

allocating an initial value to a distance of phoneme-by-phoneme; and

performing a phoneme recognition using a training database.

4. The method of claim 3, wherein the performing of the phoneme recognition comprises:

performing a dynamic matching with respect to a result of the phoneme recognition and a phoneme string corresponding to vocabularies of the training database;

estimating an optimum matching pair by back tracking; and

estimating a number of matchings of the phoneme-by-phoneme and updating the distance.

5. The method of claim 1, wherein estimating the phoneme confusion matrix comprises:

estimating a continuous HMM or a semi-continuous HMM for each phoneme by using a training database; and

estimating a distance of phoneme-by-phoneme.

6. The method of claim 5, wherein the estimating of the distance comprises:

utilizing a Bhattacharya distance, in the case of the continuous HMM; and

estimating an amount of information loss, in the case of the semi-continuous HMM.

7. The method of claim 1, wherein the determining whether the input speech signal is in-vocabulary comprises:

determining the input speech signal as in-vocabulary, when the calculated score satisfies a set numerical value; and

determining the input speech signal as an out-of-vocabulary, when the calculated score does not satisfy the set numerical value.

8. The method of claim 1, wherein the determining whether the input speech signal is in-vocabulary comprises:

utilizing a predetermined weight for the calculated score to correct the calculated score.

9. The method of claim 7, wherein the determining the input speech signal as out-of-vocabulary comprises:

performing a rejection due to a recognition error with respect to the input speech signal.

10. A non-transitory computer readable storage medium storing a program for implementing a recognition confidence measurement method comprising:

detecting a feature vector of an input speech signal through normalization of a histogram and extracting a phoneme string from the feature vector of the input speech signal;

obtaining information about a distance between a phoneme string of the input speech signal and a phoneme string of a predetermined vocabulary, by a predetermined phoneme confusion matrix which sets the distance to be decreased in proportion to the increase of a matching degree between phonemes;

extracting as candidates, the phoneme string of the vocabulary that has a higher similarity and that has a comparatively shorter distance from the speech signal, from a predetermined dictionary;

estimating a lexical distance between the extracted candidates, the estimating of the lexical distance comprising performing a dynamic matching of a pair of candidates selected from the extracted candidates; and

determining whether the input speech signal is an in-vocabulary, based on the lexical distance, wherein the dynamic matching comprises;

for each phoneme in a first candidate of the selected pair of candidates, determining as a first matching pair the phoneme in the first candidate and the corresponding phoneme in a second candidate of the selected pair of candidates and determining as a second matching pair the phoneme in the first candidate and a phoneme having an identical shape to the phoneme in the first candidate;

wherein the estimating of the lexical distance further comprises;

calculating a score for the pair of candidates; and

estimating the lexical distance using the calculated score;

wherein the calculating of the score calculates the score using the phoneme confusion matrix.

11. A recognition confidence measurement system comprising:

at least one processor to control one or more of the following units;

a phoneme string extraction unit detecting a feature vector of an input speech signal through normalization of a histogram and extracting a phoneme string from the feature vector of the input speech signal;

a candidate extraction unit obtaining information about a distance between a phoneme string of the input speech signal and a phoneme string of a predetermined vocabulary, by a predetermined phoneme confusion matrix which sets the distance to be decreased in proportion to the increase of a matching degree between phonemes, extracting as candidates the phoneme string of the vocabulary that has a higher similarity and that has a comparatively shorter distance from the speech signal, from a predetermined dictionary;

a distance estimation unit estimating a lexical distance between the extracted candidates, the estimating of the lexical distance comprising performing a dynamic matching of a pair of candidates selected from the extracted candidates; and

a registration determination unit determining whether the input speech signal is an in-vocabulary, based on the lexical distance, wherein the dynamic matching comprises;

for each phoneme in a first candidate of the selected pair of candidates, determining as a first matching pair the phoneme in the first candidate and the corresponding phoneme in a second candidate of the selected pair of candidates and determining as a second matching pair the phoneme in the first candidate and a phoneme having an identical shape to the phoneme in the first candidate;

wherein the estimating of the lexical distance further comprises;

calculating a score for the pair of candidates; and

estimating the lexical distance using the calculated score;

wherein the calculating of the score calculates the score using the phoneme confusion matrix.

12. The system of claim 11, wherein the phoneme string extraction unit extracts an optimum phoneme string according to each language by using a Hidden Markov Model (HMM) and a predetermined phoneme grammar for the each language.

13. The system of claim 11, wherein the registration determination unit determines the input speech signal as in-vocabulary, when the calculated score satisfies a set numerical value and determines the input speech signal as an out-of-vocabulary, when the calculated score does not satisfy the set numerical value.

14. The system of claim 11, wherein the registration determination unit utilizes a predetermined weight for the calculated score to correct the calculated score.

15. The system of claim 11, wherein the registration determination unit performs a rejection due to a recognition error with respect to the input speech signal.

16. A recognition confidence measurement method comprising:

extracting candidates by matching a phoneme string of a speech signal, extracted by a feature vector which is detected from the speech signal through normalization of a histogram, and phoneme strings of vocabularies registered in a predetermined dictionary;

obtaining information about a distance between a phoneme string of the input speech signal and a phoneme string of a predetermined vocabulary, by a predetermined phoneme confusion matrix which sets the distance to be decreased in proportion to the increase of a matching degree between phonemes;

extracting as candidates, the phoneme string of the vocabulary that has a higher similarity and that has a comparatively shorter distance from the speech signal, from a predetermined dictionary;

estimating a lexical distance between the extracted candidates, the estimating of the lexical distance comprising performing a dynamic matching of a pair of candidates selected from the extracted candidates; and

determining whether the speech signal is an in-vocabulary, based on the lexical distance, wherein the dynamic matching comprises;

for each phoneme in a first candidate of the selected pair of candidates, determining as a first matching pair the phoneme in the first candidate and the corresponding phoneme in a second candidate of the selected pair of candidates and determining as a second matching pair the phoneme in the first candidate and a phoneme having an identical shape to the phoneme in the first candidate;

wherein the estimating of the lexical distance further comprises;

calculating a score for the pair of candidates; and

estimating the lexical distance using the calculated score;

wherein the calculating of the score calculates the score using the phoneme confusion matrix.

17. A non-transitory medium comprising computer readable instructions implementing the method of claim 16.