Speech recognition method

• US 5,050,215 A
• Filed: 05/10/1990
• Issued: 09/17/1991
• Est. Priority Date: 10/12/1987
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Nishimura, Masafumi
• Primary Examiner(s): Kemeny, Emanuel S.
• Assistant Examiner(s): Knepper, David D.
• Application Number:
  US07/524,689
• Time in Patent Office: 495 Days
• Field of Search: 381/41-45, 364/513.5
• US Class Current: 704/256.4
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence

Abstract

For circumstance adaption, for example, speaker adaption, confusion coefficients between the labels of the label alphabet for initial training and those for adaption are determined by alignment of adaption speech with the corresponding initially trained Markov model. That is, each piece of adaptation speech is aligned with a corresponding initially trained Markov model by the Viterbi algorithm, and each label in the adaption speech is mapped onto one of the states of the Markov models. In respect of each adaptation lable ID, the parameter values for each initial training label of the states which are mapped onto the adaptation label in concern are accumulated and normalized to generate a confusion coefficient between each initial training label and each adaptation label. The parameter table of each Markov model is rewritten in respect of the adaptation label alphabet using the confusion coefficients.

Claims

1. In a speech recognition method wherein Markov models trained by an initial training label set and initial training speech are adapted using adaptation speech, an improvement comprising:

interpreting said adaptation speech into adaptation label strings using an adaptation label set different from said initial training label set;

connecting each label in each of said adaptation label strings with each state or each state transition of a Markov model which corresponds to the adaptation label strings concerned;

determining a confusion probability of each label in said initial training label set and each label in said adaptation label set being confused with each other, based on connection between each label in said adaptation label set and each of said states or state transitions, and parameter values of the Markov model concerned in respect of said initial training set; and

determining parameter values of each of said Markov models in respect of said adaptation label set, based on said confusion probabilities and said parameter values of the Markov model concerned in respect of said initial label set.

2. A speech recognition method as described in claim 1, wherein a prototype of each label in said initial training label set is modified using said adaptation speech to generate a prototype of each label in said adaptation label set.

3. A speech recognition method as described in claim 2, wherein feature vectors extracted from said adaptation speech are classified into classes according to label prototypes of said initial training label set, and an average of each of said classes is used as a corresponding prototype of said adaptation label set.

4. A speech recognition method as described in claim 3 wherein said connection between each of said label and each of said states or state transitions is established by a path by which each of said adaptation label strings is linearly aligned with a Markov model corresponding to the adaptation label string concerned.

5. A speech recognition method comprising the steps of:

providing a set of Markov models, each Markov model corresponding to an element of speech information, each Markov model having states, a set of initial labels, and transitions between states, each label having an initial conditional probability for each state or transition;

converting an element of adaptation speech information into a string of adaptation labels using an adaptation label set different from the initial label set;

identifying a corresponding Markov model from the set of Markov models which corresponds to the string of adaptation labels;

estimating for the corresponding Markov model an optimum string of states and transitions corresponding to the string of adaptation labels;

calculating for each adaptation label Lj the probabilities P(Lj|Li) of confusing the adaptation label Lj with each initial label Li, said probabilities being calculated based on the initial conditional probabilities of the initial labels Li for each of the states or transitions corresponding to the adaptation label Lj;

calculating for each adaptation label Lj a revised conditional probability P(Lj|Sk) of occurrence of the adaptation label Lj for a first state or transition Sk based on the confusion probabilities P(Lj|Li) and based on the initial conditional probabilities P(Li|Sk) of the initial labels Li for the first state or transition; and

updating the corresponding Markov model by replacing the initial conditional probabilities of the initial labels with the revised conditional probabilities of the adaptation labels.

6. A method as claimed in claim 5, characterized in that the step of calculating the probabilities P(Lj|Li) of confusing the adaptation labels Lj with the label Li comprises the steps of:

summing for each Lj the probabilities P(Li|Sk) for all states or transitions which corresponds to the adaptation label Lj to produce a count C(Lj,Li) for each Li and for each Lj;

summing for each Li the count C(Lj,Li) for all Lj to produce a divisor Di for each Li; and

dividing for each Li and Lj the count C(Lj,Li) by the divisor Di to produce a quotient equal to P(Lj|Li) for each Lj.

7. A method as claimed in claim 6, characterized in that the step of calculating a revised conditional probability P(Lj|Sk) comprises the steps of multiplying P(Lj|Li) by P(Li|Sk) for each Li to produce products Qi, and then summing the products Qi to produce a total equal to P(Lj|Sk).

8. A speech recognition method comprising the steps of:

providing a set of Markov models, each Markov model corresponding to an element of speech information, each Markov model having states, transitions between states, probabilities of making transitions and probabilities of outputting labels at each of the transitions or states, said probabilities initially trained with a initial training label set;

converting an element of adaptation speech information into a string of adaptation labels using an adaptation label set different from the initial training label set;

identifying a corresponding Markov model from the set of Markov models which corresponds to the string of adaptation labels;

estimating for the corresponding Markov model an optimum string of states or transitions corresponding to the string of adaptation labels;

calculating a frequency of each of said adaptation labels corresponding to each of said states or transitions by said estimating;

calculating a probability of each of said initial training labels and each of said adaptation labels being confused with each other based on the frequency of the adaptation label concerned corresponding to each of said states or transitions and the probability of outputting the initial training label concerned at each of said states or transitions;

calculating a revised probability of outputting each of said adaptation labels at each of said states or transitions based on the confusion probabilities and the initial probabilities of outputting the initial training labels; and

updating the corresponding Markov model by replacing the probabilities of outputting the initial training labels at each of said states or transitions with the revised probabilities of outputting the adaptation labels at each of said states or transitions.

9. A speech recognition method comprising the steps of:

providing a set of Markov models, each Markov model corresponding to an element of speech information, each Markov model having states, transitions between states, probabilities of making transitions, and probabilities of outputting initial labels from a set of initial labels at each of the transitions or states, each initial label representing a range of values of at least one measurable feature of a portion of an utterance;

changing the range of values represented by at least one label in the set of initial labels to produce a set of adaptation labels;

converting an element of adaptation speech information into a string of adaptation labels using the set of adaptation labels;

identifying a corresponding Markov model from the set of Markov models which corresponds to the string of adaptation labels;

estimating for the corresponding Markov model an optimum string of states and transitions corresponding to the string of adaptation labels;

calculating for each adaptation label Lj the probabilities P(Lj|Li) of confusing the adaptation label Lj with each initial label Li, said probabilities being calculated based on the initial conditional probabilities of the initial labels Li for each of the states or transitions corresponding to the adaptation label Lj;

calculating for each adaptation label Lj a revised conditional probability P(Lj|Sk) of occurrence of the adaptation label Lj for a first state or transition Sk based on the confusion probabilities P(Lj|Li) and based on the initial conditional probabilities P(Li|Sk) of the initial labels Li for the first state or transition; and

updating the corresponding Markov model by replacing the initial conditional probabilities of the initial labels with the revised conditional probabilities of the adaptation labels.

10. A method as claimed in claim 9, characterized in that the step of calculating the probabilities P(Lj|Li) of confusing the adaptation labels Lj with the label Li comprises the steps of:

summing for each Lj the probabilities P(Li|Sk) for all states or transitions which corresponds to the adaptation label Lj to produce a count C(Lj,Li) for each Li and for each Lj;

summing for each Li the count C(Lj,Li) for all Lj to produce a divisor Di for each Li; and

dividing for each Li and Lj the count C(Lj,Li) by the divisor Di to produce a quotient equal to P(Lj|Li) for each Lj.

11. A method as claimed in claim 10, characterized in that the step of calculating a revised conditional probability P(Lj|Sk) comprises the steps of multiplying P(Lj|Li) by P(Li|Sk) for each Li to produce products Qi, and then summing the products Qi to produce a total equal to P(Lj|Sk).