Speaker-trained speech recognizer having the capability of detecting confusingly similar vocabulary words

• US 4,972,485 A
• Filed: 05/23/1989
• Issued: 11/20/1990
• Est. Priority Date: 03/25/1986
• Status: Expired due to Term

• Current Assignee: AT&T, Inc.
• Original Assignee: AT&T, Inc.
• Inventors: Roe, David B., Dautrich, Bruce A., Goeddel, Thomas W.
• Primary Examiner(s): Shaw, Dale M.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US07/356,589
• Time in Patent Office: 546 Days
• Field of Search: 364/513.5, 381/41-45, 382/13-15
• US Class Current: 704/251
• CPC Class Codes: G10L 15/07 to the speaker

Abstract

During a training sequence, a speaker-trained speech recognizer detects and signals the speaker when vocabulary word pairs are potentially confusing to the recognizer. Each vocabulary word is converted into feature signals and then parameters representing a predetermined reference model of that word. The feature signals of a subsequent potential vocabulary word are compared against the reference model of each vocabulary word previously stored in the recognizer memory. The speaker is signaled when the potential vocabulary word is confusingly similar to one of the existing vocabulary words.

Claims

1. A speaker-trained speech recognizer for selecting words to be added to a memory of said recognizer, said recognizer comprising

means for extracting a plurality of feature signals from a received word utterance from a speaker,

means for generating a plurality of parameters derived from said plurality of feature signals,

means for comparing a plurality of feature signals of a first word utterance extracted by said extracting means against a plurality of stored parameters of a different previously received word utterance using a predetermined criteria, an output of said comparing means being determined by the difference between said plurality of feature signals of said first word utterance and the plurality of stored parameters of said previously received word utterance,

means for signaling said speaker to utter into said recognizer a second word utterance, which is different from said first word utterance, when the output of said comparing means is less than a predetermined value indicating a similarity between said feature signals of said first word utterance and said previously received word utterance parameters, and

means for storing a plurality of parameters of said first word utterance as another different word utterance in said memory when the output of said comparing means indicates that said difference is not less than said predetermined value.

2. The speaker-trained speech recognizer of claim 1 wherein said plurality of parameters of each word utterance is a Hidden Markov Model thereof.

3. The speaker-trained speech recognizer of claim 1 wherein said extracting means includes a filter bank means.

4. The speaker-trained speech recognizer of claim 1 wherein said comparing means includes a Viterbi scoring means.

5. The speaker-trained speech recognizer of claim 1 wherein

said storing means stores the plurality of parameters of a group of previously received different word utterances in said memory, and wherein

said comparing means includes

previous word scoring means for scoring said plurality of feature signals of said first word utterance against the plurality of parameters of each previously received different word utterance of said group using said predetermined criteria, and

wherein said signaling means signals said speaker when a score of any of the previously received different word utterance in said group is less than said predetermined value.

6. The speaker-trained speech recognizer of claim 5 wherein said comparing means further includes

first word scoring means for scoring said plurality of feature signals of said first word utterance against the plurality of parameters of said first word utterance derived using said generating means, and wherein

said comparing means utilizes an output of said first word scoring means and outputs of said previous word scoring means to determine said predetermined value.

7. The speaker-trained speech recognizer of claim 6 wherein

a lowest score from said previous word scoring means is S(MIN),

said output of said present word scoring means is S(N+1), and

said predetermined value is equal to S(N+1)-S(MIN).

8. The speaker-trained speech recognizer of claim 6 wherein

a lowest score from said previous word scoring means is S(MIN),

said output of said present word scoring means is S(N+1), and

said predetermined value is equal to S(N+1) divided by S(MIN).

9. The speaker-trained speech recognizer of claim 6 wherein

a lowest score from said previous word scoring means is S(MIN) the duration-normalized version of S(MIN), and

said predetermined value is equal to S(MIN).

10. The speaker-trained speech recognizer of claim 1 including

means for updating said stored plurality of parameters of said first word utterance in said memory in response to a received repetition of said first word utterance from said speaker.

11. A method of operating a speaker-trained speech recognizer comprising the steps of

extracting a plurality of feature signals from a received word utterance from a speaker,

generating a plurality of parameters derived from said plurality of feature signals,

comparing a plurality of feature signals of a first word utterance extracted by said extracting step against a plurality of stored parameters of a different previously received word utterance using a predetermined criteria, an output of said comparing being determined by the difference between said plurality of feature signals of said first word utterance and the plurality of stored parameters of said previously received different word utterance,

signaling said speaker to utter into said recognizer a second word utterance, which is different from said first word utterance, when the output of said comparing step is less than a predetermined value indicating a similarity between said feature signals of said first word utterance and said previously received word utterance parameters, and

storing a plurality of parameters of said first word utterance as another different word utterance in said memory when the output of said comparing step indicates that said difference is not less than said predetermined value.