Pattern recognition system

• US 4,955,056 A
• Filed: 07/16/1986
• Issued: 09/04/1990
• Est. Priority Date: 07/16/1985
• Status: Expired due to Term

• Current Assignee: British Telecommunications PLC (BT Group PLC)
• Original Assignee: British Telecommunications (BT Group Plc) (BT Group PLC)
• Inventors: Stentiford, Frederick W. M.
• Primary Examiner(s): Harkcom, Gary V.
• Assistant Examiner(s): Merecki, John A.
• Application Number:
  US06/886,072
• Time in Patent Office: 1,511 Days
• Field of Search: 381/41-45, 364/513.5, 382/10-11, 382/16, 382/18, 382/19, 382/30
• US Class Current: 704/239
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence

Abstract

A waveform recognition system including a plurality of detectors of features having the combined presence at a plurality of instants spaced at predetermined intervals relative to each other in time of instantaneous amplitudes each satisfying respectively predetermined constraints; apparatus for assigning a plurality of labels and corresponding confidence measures to each of successive portions of the waveform in dependence on the features detected in the portions and storing each label in a buffer corresponding to the rank of the confidence with which the label is assigned relative to other labels assigned to the same portion of data and apparatus for outputting labels from that buffer containing labels assigned with the highest confidence whose confidence measures are in a predetermined relationship with those of adjacent labels in the same buffer when the confidence measures of labels in other buffers containing labels assigned with confidence measures of lower rank satisfy predetermined conditions.

Claims

1. Speech recognition apparatus comprising:

(a) input means for receiving, in successive overlapping temporal portions, an electrical signal containing speech data;

(b) a feature detection device responsive to said electrical signal over a temporal portion thereof for detecting the presence of a plurality of predetermined features within said portion; and

(c) decision means for indicating recognition of elements of speech, each said element corresponding to the presence of a predetermined combination of said detected features, the said decision means including;

(i) assignment means for assigning a label corresponding to one of said elements of speech to each said portion in dependence on the features detected therein, together with a corresponding confidence measure indicating the degree of confidence in the correctness of the assignment of that label;

(ii) an output buffer connected to said assignment means for storing values corresponding to a plurality of said successive portions forming a temporal array, said values comprising, for each said portion, timing information defining the relative position in time of that portion, and the label and corresponding confidence measure assigned to that portion; and

(iii) output means for indicating recognition of an element of speech, by outputting from said output buffer the labels and timing information for those portions in said array whose corresponding successive confidence measures define local maxima in said array.

2. A speech recognition apparatus according to claim 1 wherein:

said assignment means includes means for assigning a plurality of different labels from a reference set of labels and corresponding confidence measures indicating a degree of confidence in the correct assignment of each label to each of said successive portions of data in dependence on the features detected in said portions by said feature detection device;

said decision means further comprises a plurality of other lower rank buffer means, each of successively lower rank for storing different respective values corresponding to said plurality of said successive portions of data and forming for each rank a one dimensional array for each portion of data, having (i) timing information defining the temporal position of the portion relative to others of said portions, (ii) one of said labels and (iii) a corresponding confidence measure assigned to that portion by said assignment means;

said output buffer means contains values having confidence measures which indicate the highest confidence in the correct assignment of the corresponding label of all labels in said reference set in respect of the corresponding portion of data; and

said output means outputs electronic recognition signals when the confidence measures of labels in the others of said lower rank buffer means are not rising with respect to time.

5. Apparatus according to claim 1, 2, 3 or 4 wherein said output means operates to indicate recognition if the label to be output has a corresponding confidence measure indicating a greater degree of confidence in its correctness of assignment than that of any of the labels corresponding to a predetermined number of succeeding portions in said array.

6. Apparatus according to claim 1, 2, 3 or 4 wherein:

said buffer means includes one or more shift registers,

said decision means tests the confidence measures of said labels in said buffer means at the input to each buffer means, and

said output means outputs labels and timing information if, before the label to be output has reached the output, no further confidence measure maxima greater than that to be output have been detected at the input of an output buffer means.

7. Apparatus according to claim 1, 2, 3 or 4 wherein said array contains only values for portions of data spanning a total time length less than the time duration of said element to be recognized.

8. Apparatus according to claim 1, 2, 3 or 4 wherein:

(a) said input means includes an input buffer shift register comprising a series of cells through which said signal can be continuously stepped, the contents of the buffer constituting a said temporal portion; and

(b) said decision means comprises means for reading the feature detection device at steps of said signal through said register and comparing the readings for said features with predetermined reference vectors each having a corresponding label, and assigning that label whose reference vector most closely matches said readings together with a corresponding confidence measure to the portion of said signal which produced said readings.

9. Apparatus according to claim 8 wherein said input buffer is dimensioned to contain a signal portion corresponding in time duration to at least the length of the longest element of speech to be recognized, and said labels are assigned at each step of said signal through said input buffer.

10. Apparatus according to claim 8 wherein said feature detection device further comprises means including a cumulative store for each detectable feature, the content of said stores indicating whether a feature has been detected since the store was last cleared and means for comparing the contents of the stores with said reference vectors at each step of said signal through said input buffer, said cumulative stores being cleared after a predetermined time.

11. Apparatus according to claim 10 wherein said cumulative stores are cleared after a label is output from said system.

12. Character string matching apparatus comprising apparatus according to claim 8 wherein said input buffer is connected to receive a binary coded character string, and each said cell stores a binary-coded character, and each said feature corresponds to at least one binary coded character.

3. Pattern recognition apparatus comprising:

(a) means for receiving an input electrical signal;

(b) decision means for indicating recognition of reference pattern elements by outputting corresponding reference labels for portions of said signal, said decision means including assignment means for assigning, to each of said portions, a plurality of said reference labels together with corresponding confidence measures indicating the degree of confidence in the correctness of assignment of each such label;

(c) a plurality of buffer means, each for storing values corresponding to a plurality of successive said portions and forming a temporal array wherein said values include, for each said portion, (i) timing information defining the relative position in time of the portion, (ii) one of said labels and (iii) the corresponding confidence measure, and each buffer means containing labels having corresponding confidence measures, one of said buffer means being an output buffer means containing values whose corresponding confidence measures indicate the highest confidence in the correctness of assignment of the corresponding label of all labels in said reference labels in respect of the corresponding said portion; and

(d) output means for indicating recognition of a pattern, by outputting from said output buffer means, labels and timing information corresponding to those portions in said array whose successive confidence measures define local maxima in said array, when the successive confidence measures of labels in the others of said buffer means are not rising with time.

4. Apparatus according to claim 3 wherein said output means operates to delay indicating recognition and to store label and timing information for a predetermined time, and to replace said stored label and timing information with those for any said element subsequently recognized during said predetermined time having a confidence measure indicating a greater confidence in the correctness of assignment of the corresponding label than that of said stored label, and to output the stored label and timing information at the end of said predetermined time.

13. Speech recognition apparatus comprising pattern recognition apparatus according to 3 or 4.

14. A method of detecting the occurrence of speech events in a speech signal comprising the steps of:

(a) partitioning the speech signal into successive, overlapping temporal portions S_i ;

(b) comparing each portion S_i with a vocabulary of speech events and generating, for each said speech event, a measure C_i of the similarity between the portion S_i and that speech event L_n ;

(c) for each portion S_i, finding the highest-ranking speech event L₁, and at least the next-highest-ranking speech event L₂, ranked by their similarity to that portion in accordance with their similarity measures C₁, C₂ ;

(d) storing an indication of the highest-ranking such speech event L₁ and of the corresponding similarity measure C₁ for each of a sequence of successive portions S₁, S₂. . . , spanning a time interval at least comparable to the length of the longest speech event in the said vocabulary;

(e) locating the temporal position of the detected highest-ranking speech event L₁, within the sequence, by finding a local maximum portion S_K indicated by its highest-ranking similarity measure C₁ to be more similar to the highest-ranking speech event L₁ than those preceding it and succeeding it in the sequence; and

(f) indicating recognition of the highest-ranking speech event L₁ at the temporal position corresponding to the local maximum portion S_K.

15. A method according to claim 14 further comprising the steps of:

(g) storing also, for each portion of the said sequence S₁, S₂ . . . , an indication of at least one of the next-highest ranking speech events and corresponding indication(s) of similarity measure;

(h) upon finding a said local maximum portion S_K, detecting whether successive portions including the said local maximum portion are becoming more similar to a lower-ranking speech event; and

if so,

(i) inhibiting indication of recognition.

16. A method according to claim 15 in which the said step of inhibiting indication of recognition comprises the step of:

(a) storing data corresponding to the indication which would otherwise have been made;

(b) determining whether, in a predetermined time following the said local maximum, there is a further local maximum portion, and, if so,

(c) determining which of the stored local maximum portion S_K and any further such local maximum portions was more similar to its corresponding speech event, and

(d) indicating recognition of that corresponding speech event.