Method for recognizing patterns in time-variant measurement signals

• US 5,555,344 A
• Filed: 03/18/1994
• Issued: 09/10/1996
• Est. Priority Date: 09/20/1991
• Status: Expired due to Term

• Current Assignee: Unify GmbH & Co. KG (Atos SE)
• Original Assignee: Siemens AG
• Inventors: Zunkler, Klaus
• Primary Examiner(s): Downs, Robert W.
• Assistant Examiner(s): HAFIZ, TARIQ R
• Application Number:
  US08/211,201
• Time in Patent Office: 907 Days
• Field of Search: 381/41-43, 395/2, 395/2.4, 395/2.45, 395/2.51, 395/2.5, 395/2.52, 395/2.65, 395/2.49, 395/2.63-2.64
• US Class Current: 704/242
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

In automatic speech recognition, confusion easily arises between phonetically similar words (for example, the German words "zwei" and "drei") in the case of previous recognition systems. Confusion of words which differ only in a single phoneme (for example, German phonemes "dem" and "den") occurs particularly easily with these recognition systems. In order to solve this problem, a method for recognizing patterns in time-variant measurement signals is specified which permits an improved discrimination between such signals by reclassifying in pairs. This method combines the Viterbi decoding algorithm with the method of hidden Markov models, the discrimination-relevant features being examined separately in a second step after the main classification. In this case, different components of feature vectors are weighted differently, it being the case that by contrast with known approaches these weightings are performed in a theoretically based way. The method is suitable, inter alia, for improving speech-recognizing systems.

Claims

1. A method for recognizing patterns in time-variant measurement signals by classifying a temporal sequence of feature vectors and reclassification in pairs, comprising the steps of:

segmenting the sequence of feature vectors which is to be classified using a Viterbi decoding algorithm, this sequence to be classified being compared with a set of hidden Markov models;

calculating for each hidden Markov model a total emission probability for the generation of the sequence to be classified by this hidden Markov model;

determining an optimum assignment path from feature vectors to states of the hidden Markov models by backtracking;

calculating, for at least one pair of hidden Markov models, modified total emission probabilities for each hidden Markov model of said at least one pair on a precondition that a respective other hidden Markov model of a same pair competes with the hidden Markov model under review, the total emission probability being calculated, for generating the sequence to be classified by a hidden Markov model, by calculating for all feature vectors of the sequence to be classified and for all states of the hidden Markov model a local logarithmic emission probability for generating the respective feature vector by the respective state, and by calculating an accumulated logarithmic emission probability for each state as a sum of its local logarithmic emission probability and an accumulated logarithmic emission probability of its best possible predecessor state, the best possible predecessor state being logged;

determining a respective more probable hidden Markov model of said at least one pair;

selecting the hidden Markov model having the highest total emission probability from among all pairs under review.

2. The method as claimed in claim 1, wherein the method further comprises carrying out a discriminating comparison of the assignments of two hidden Markov models by calculating for all feature vectors of the temporal sequence a local logarithmic modified emission probability for generating the respective feature vector by the corresponding state of the respective hidden Markov model by adding these local logarithmic modified emission probabilities recursively along the already calculated assignment path to an accumulated logarithmic modified emission probability.

3. The method as claimed in claim 2, wherein the method further comprises calculating the modified local emission probability, for generating a feature vector by a state of a hidden Markov model, by calculating a logarithmic emission probability for each component of this feature vector and multiplying the logarithmic emission probability by a weighting, and by summing these weighted logarithmic emission probabilities over all components of this feature vector;

the weighting of a component of a feature vector being in this case a measure of the reliability of this component on an assumption that only two specific states of a hidden Markov model are to be compared with one another.

4. The method as claimed in claim 3, wherein the weightings are determined using a statistic quality measure for the reliability of a feature vector component.

5. The method as claimed in claim 3, wherein the weightings are determined using a training method.

6. A method for recognizing patterns in time-variant measurement signals by classifying a temporal sequence of feature vectors and reclassification in pairs, comprising the steps of:

segmenting the sequence of feature vectors which is to be classified using a Viterbi decoding algorithm, this sequence to be classified being compared with a set of hidden Markov models;

calculating, for pairs of hidden Markov models, modified total emission probabilities for each hidden Markov model of a respective pair of the hidden Markov models on a precondition that a respective other hidden Markov model of a same pair competes with the hidden Markov model under review, the total emission probability being calculated, for generating the sequence to be classified by a hidden Markov model, by calculating for all feature vectors of the sequence to be classified and for all states of the hidden Markov model a local logarithmic emission probability for generating the respective feature vector by the respective state, and by calculating an accumulated logarithmic emission probability for each state as a sum of its local logarithmic emission probability and an accumulated logarithmic emission probability of its best possible predecessor state, the best possible predecessor state being logged;

determining a respective more probable hidden Markov model of said respective pair;

selecting the hidden Markov model having the highest total emission probability from among all pairs under review; and

determining an optimum assignment path from feature vectors to states of the hidden Markov models by backtracking.

7. The method as claimed in claim 6, wherein the method further comprises carrying out the discriminating comparison of the assignments of two hidden Markov models by calculating for all feature vectors of the temporal sequence a local logarithmic modified emission probability for generating the respective feature vector by the corresponding state of the respective hidden Markov model by adding these local logarithmic modified emission probabilities recursively along the already calculated assignment path to an accumulated logarithmic modified emission probability.

8. The method as claimed in claim 7, wherein the method further comprises calculating the modified local emission probability, for generating a feature vector by a state of a hidden Markov model, by calculating a logarithmic emission probability for each component of this feature vector and multiplying the logarithmic emission probability by a weighting, and by summing these weighted logarithmic emission probabilities over all components of this feature vector;

the weighting of a component of a feature vector being in this case a measure of the reliability of this component on an assumption that only two specific states of a hidden Markov model are to be compared with one another.

9. The method as claimed in claim 8, wherein the weightings are determined using a statistic quality measure for the reliability of a feature vector component.

10. The method as claimed in claim 8, wherein the weightings are determined using a training method.

11. A method for recognizing patterns in time-variant measurement signals by classifying a temporal sequence of feature vectors and reclassification in pairs, comprising the steps of:

segmenting the sequence of feature vectors which is to be classified using a Viterbi decoding algorithm, this sequence to be classified being compared with a set of hidden Markov models;

calculating for each hidden Markov model a total emission probability for the generation of the sequence to be classified by this hidden Markov model;

determining an optimum assignment path from feature vectors to states of the hidden Markov models by backtracking;

calculating, for at least one pair of hidden Markov models, modified total emission probabilities for each hidden Markov model of said at least one pair on a precondition that a respective other hidden Markov model of a same pair competes with the hidden Markov model under review, a local logarithmic modified emission probability being calculated for all feature vectors of the temporal sequence for generating a respective feature vector by the corresponding state of the respective hidden Markov model by adding local logarithmic modified emission probabilities recursively along the already calculated assignment path to an accumulated logarithmic modified emission probability, a total emission probability being calculated, for generating the sequence to be classified by a hidden Markov model, by calculating for all feature vectors of the sequence to be classified and for all states of the hidden Markov model a local logarithmic emission probability for generating the respective feature vector by the respective state, and by calculating an accumulated logarithmic emission probability for each state as a sum of its local logarithmic emission probability and an accumulated logarithmic emission probability of its best possible predecessor state, the best possible predecessor state being logged;

determining a respective more probable hidden Markov model of said at least one pair; and

selecting the hidden Markov model having the highest total emission probability from among all pairs under review.

12. The method as claimed in claim 11, wherein the method further comprises calculating the modified local emission probability, for generating a feature vector by a state of a hidden Markov model, by calculating a logarithmic emission probability for each component of this feature vector and multiplying the logarithmic emission probability by a weighting, and by summing these weighted logarithmic emission probabilities over all components of this feature vector;

the weighting of a component of a feature vector being in this case a measure of the reliability of this component on an assumption that only two specific states of a hidden Markov model are to be compared with one another.

13. The method as claimed in claim 12, wherein the weightings are determined using a statistic quality measure for the reliability of a feature vector component.

14. The method as claimed in claim 12, wherein the weightings are determined using a training method.