Speech recognizer

• US 6,868,382 B2
• Filed: 03/09/2001
• Issued: 03/15/2005
• Est. Priority Date: 09/09/1998
• Status: Expired due to Fees

• Current Assignee: Asahi Kasei Kabushiki Kaisha (Asahi Kasei Corporation)
• Original Assignee: Asahi Kasei Kabushiki Kaisha (Asahi Kasei Corporation)
• Inventors: Shozakai, Makoto
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): Nolan, Daniel A
• Application Number:
  US09/802,768
• Publication Number:
  US 20020116192A1
• Time in Patent Office: 1,467 Days
• Field of Search: 704/254, 704/256, 704/244, 704/240, 704/237, 704/222
• US Class Current: 704/254
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

The generic word label series used for recognition of words uttered by unspecified speakers are stored in the vocabulary label network accumulation processing. The speech of a particular speaker is entered. Based on the input speech, the registered word label series extraction processing generates the registered word label series. The registered word label series of the particular speaker can then be registered with the vocabulary label network accumulation processing.

Claims

1. A speech recognizer for performing word recognition on input speech by using information on models of speech units each shorter than a word, the speech recognizer comprising:

vocabulary label network accumulation means for accumulating label series of said speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers;

registered word label series extraction means for generating label series of said speech units for registered words from input speech of a particular speaker; and

registration means for storing the label series of speech units for the generic words commonly used for word recognition of input speech of said unspecified speakers and the generated registered word label series in the form of parallel networks in said vocabulary label network accumulation means;

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

2. The speech recognizer as claimed in claim 1, further comprising means for registering said generic words with said vocabulary label network accumulation means.

3. The speech recognizer for performing word recognition on input speech by using information on models of speech units each shorter than a word the speech recognizer comprising:

vocabulary label network accumulation means for accumulating label series of said speech units for generic words commonly use to perform word recognition on input speech of unspecified speakers;

registered word label series extraction means for generating label series satisfying a connection of said speech units and having the highest probability in the label series of said speech units for registered words from input speech of a particular speaker by using a network in which said connection of the speech units related to the connections of speech units is described; and

registration means for registering to add the generated registered word label series to said vocabulary label network accumulation means;

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

4. The speech recognizer as claimed in claim 3, further comprising means for registering said generic words with said vocabulary label network accumulation means.

5. The speech recognizer for performing word recognition on input speech by using information on models of speech units each shorter than a word, the speech recognizer comprising:

vocabulary label network accumulation means for accumulating label series of said speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers;

registered word label series extraction means for generating label series satisfying a connection of said speech units and having the highest probability in the label series of said speech units for registered words from input speech of a particular speaker by using a network in which said connection of the speech units related to the connections of speech units is described; and

registration means for storing said label series of speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers and the generated registered word label series in the form of parallel networks in said vocabulary label network accumulation means;

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

6. The speech recognizer as claimed in claim 5, further comprising means for registering said generic words with said vocabulary label network accumulation means.

7. A speech recognition method for performing word recognition on input speech by using information on models of speech units each shorter than a word,

wherein label series of said speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers are accumulated in vocabulary label network accumulation means;

said method comprising steps of;

generating label series of said speech units for registered words from input speech of a particular speaker; and

storing said label series of speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers and the generated registered word label series in the form of parallel networks in said vocabulary label network accumulation means

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

8. The speech recognition method as claimed in claim 7, wherein said generic words can be registered with said vocabulary label network accumulation means.

9. A speech recognition method for performing word recognition on input speech by using information on models of speech units each shorter than a word.

wherein label series of said speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers are accumulated in vocabulary label network accumulation means;

said method comprising steps of;

generated label series satisfying a connection of said speech units and having the highest probability in the label series of said speech units for registered words from input speech of a particular speaker by using a network in which said connection of the speech units related to the connections of speech units is described; and

registering to add the generated registered word label series to said vocabulary label network accumulation means;

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

10. The speech recognition method as claimed in claim 9, wherein said generic words is possible to be registered to said vocabulary label network accumulation means.

11. A speech recognition method for performing word recognition on input speech by using information on models of speech units each shorter than a word.

wherein label series of said speech units for generic words commonly used to perform word recognition on input speech of unspecified speakers are accumulated in vocabulary label network accumulation means:

said method comprising steps of;

generating label series satisfying a connection of said speech units and having the highest probability in the label series of said speech units for registered words from input speech of a particular speaker by using a network in which said connection of the speech units related to the connections of speech units is described; and

storing said label series of speech units for generic words commonly use to perform word recognition in input speech of unspecified speakers and the generated registered word label series in said vocabulary label network accumulation means;

wherein said speech units are acoustic events generated by dividing a Hidden Markov Model of phoneme into individual states while maintaining the values of a transition probability and an output probability and the number of states.

12. The speech recognition method as claimed in claim 11, wherein said generic words is possible to be registered to said vocabulary label network accumulation means.