Speech pattern recognition using pattern recognizers and classifiers

• US 5,794,190 A
• Filed: 11/09/1994
• Issued: 08/11/1998
• Est. Priority Date: 04/26/1990
• Status: Expired due to Fees

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications PLC (BT Group PLC)
• Inventors: Smyth, Samuel Gavin, Talintyre, John Edwrd, Linggard, Robert, Woodland, Philip Charles
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US08/338,158
• Time in Patent Office: 1,371 Days
• Field of Search: 395/2, 395/2.11, 395/2.41, 395/2.68, 395/21, 395/23, 395/24, 381/42, 381/43, 704/200, 704/202, 704/232, 704/259, 704/240, 704/243, 704/244, 704/239, 704/245
• US Class Current: 704/232
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 25/30 using neural networks

Abstract

An apparatus for recognizing predetermined patterns comprises an input for signal data; a pattern matching stage including a plurality of stores, each storing a representation of a pattern to be recognized; a similarity measurer for producing, for each representation, a measure of the similarity between the signal data and that representation; and a classification stage for deciding from the similarity measures which, if any, of the patterns to which the representations correspond is present in the signal data, and for indicating the presence of any such pattern. The classification stage makes its determination based on all the similarity measures after having been trained using information relating to known patterns and the similarity measurer from the pattern matching stage produced by inputting known patterns.

Claims

1. Apparatus for recognizing word patterns, comprising:

an input for signal data; and

at least two stages of recognition comprising;

a pattern matching stage which includes a plurality of stores each storing a representation of a pattern to be recognized and means for producing, for each said representation, a measure of the similarity between the signal data and that representation; and

a classification stage having a plurality of inputs, each input being connected to receive the measures of similarity produced for one of said representations,

the classification stage deciding on the basis of taking into account all of the said measures which, if any, of the patterns to be recognized to which the representations correspond is present in the signal data, and which indicates the presence of any such pattern,

the classification stage being trainable, with parameters of the classification stage derived from information related to a predetermined pattern represented by signal data input to the pattern matching stage and from the measures of similarity produced for each said representation from input of said signal data representing the predetermined pattern.

2. Apparatus as claimed in claim 1 wherein each pattern to be recognized corresponds to a predetermined non-zero function of at least two of a predetermined plurality of the said measures.

3. Apparatus according to claim 1 in which the classification stage receives all the said measures, and calculates for each pattern to be recognized, a corresponding predetermined Gaussian function of the measures.

4. Apparatus according to claim 1 in which the classification stage calculates a plurality of sums of the products of each of a predetermined plurality of measures and derives respective corresponding weighting constants.

5. Apparatus according to claim 4 in which the classification stage processes each said sum by a non linear compression function, and calculates for each pattern to be recognized a further sum of the products of the processed sums and derives corresponding weighting factors.

6. Apparatus according to claim 1 in which the classification stage comprises a multi-layer perceptron.

7. Apparatus according to claim 1 in which the means for producing a measure of the similarity are arranged to do so substantially independently of the duration and/or position of events in the signal data.

8. Apparatus according to claim 7 in which the means for producing a measure of the similarity comprises a Dynamic Warp calculation.

9. Apparatus according to claim 7 in which the means for producing each measure of the similarity comprises calculating the similarity of the signal data to a stored Hidden Markov state transition Model.

10. An apparatus as claimed in claim 1 wherein said signal data comprises speech signal data.

11. Apparatus for recognizing word patterns according to claim 1, wherein the pattern matching stage is separately trained on a first set of test data and the classification stage is separately trained on a second set of test data not identical to the first.

12. A speech recognizer comprising:

means to pre-process speech signals to extract significant features of those signals;

store means to hold reference features corresponding to known speech patterns to be recognized;

comparison means to compare the features extracted from said speech signals with stored reference features and to produce for each comparison a measure of the similarity therebetween; and

decision means having a plurality of inputs, each input connected to receive the measures of similarity produced for one of said features,

said decision means determining on the basis of all said measures of similarity the speech pattern to which an input signal is most similar;

characterized in that said decision means is one which has been trained using information relating to a known predetermined speech pattern represented by speech signals input to the means to pre-process speech signals and the measures of similarity produced by the comparison means for each said feature from the input of speech signals representing the said known predetermined speech pattern.

13. A recognizer as in claim 12 wherein the decision means comprises a multi-layer perceptron.

14. A recognizer as in claim 12 wherein said means to pre-process is arranged to produce mel-frequency cepstral coefficients.

15. A method of training a pattern recognition apparatus comprising:

receiving input signal data;

comparing the input signal data with each of a plurality of stored representations each corresponding to a pattern to be recognized to produce a similarity measure for each of the plurality of stored representations; and

deriving parameters for a decision means for having a plurality of inputs connected to receive said similarity measures produced for one of said representations from information relating to a predetermined pattern and from (b) the said similarity measures produced for each representation on receiving input signal data corresponding to the predetermined pattern.

16. A method according to claim 15 in which the step of deriving said parameters includes the step of calculating a Gaussian function of said measures.

17. A method according to claim 15 in which the step of deriving said parameters includes the step of calculating a sum of the products of each of the measures and corresponding weighting constants.

18. A method according to claim 17, wherein the step of calculating a sum further comprises the steps of processing each such sum using a non linear sigmoidal function, and calculating a plurality of sums of the products of the processed sums and corresponding weighting constants.

19. A method according to claim 15, in which the step of comparing the input signal with each representation is performed so as to be substantially independent of the duration and/or position of pattern events in the signal data.

20. A method according to claim 19 in which the step of comparing the input signal data with each representation comprises employing a Dynamic Warp calculation.

21. A method according to claim 20 in which the step of comparing the input signal data with each representation comprises employing a Hidden Markov Model.

22. A method according to claim 15 wherein said input signal data comprises speech signal data.

23. A trained pattern recognizer comprising apparatus trained according to the method of claim 15.