Training of markov models used in a speech recognition system
First Claim
1. In a system for decoding a vocabulary word from outputs selected from an alphabet of outputs in response to a communicated word input wherein each word in the vocabulary is represented by a baseform of at least one probabilistic finite state model and wherein each probabilistic model has transition probability items and output probability items and wherein a probability value is stored for each of at least some probability items, a method of determining probability values comprising the step of:
- biassing at least some of the stored probability values to enhance the likelihood that outputs generated in response to communication of a known word input are produced by the baseform for the known word relative to the respective likelihood of the generated outputs being produced by the baseform for at least one other word.
1 Assignment
0 Petitions
Accused Products
Abstract
In a word, or speech, recognition system for decoding a vocabulary word from outputs selected from an alphabet of outputs in response to a communicated word input wherein each word in the vocabulary is represented by a baseform of at least one probabilistic finite state model and wherein each probabilistic model has transition probability items and output probability items and wherein a value is stored for each of at least some probability items, the present invention relates to apparatus and method for determining probability values for probability items by biassing at least some of the stored values to enhance the likelihood that outputs generated in response to communication of a known word input are produced by the baseform for the known word relative to the respective likelihood of the generated outputs being produced by the baseform for at least one other word. Specifically, the current values of counts --from which probability items are derived--are adjusted by uttering a known word and determining how often probability events occur relative to (a) the model corresponding to the known uttered "correct" word and (b) the model of at least one other "incorrect" word. The current count values are increased based on the event occurrences relating to the correct word and are reduced based on the event occurrences relating to the incorrect word or words.
-
Citations
21 Claims
-
1. In a system for decoding a vocabulary word from outputs selected from an alphabet of outputs in response to a communicated word input wherein each word in the vocabulary is represented by a baseform of at least one probabilistic finite state model and wherein each probabilistic model has transition probability items and output probability items and wherein a probability value is stored for each of at least some probability items, a method of determining probability values comprising the step of:
biassing at least some of the stored probability values to enhance the likelihood that outputs generated in response to communication of a known word input are produced by the baseform for the known word relative to the respective likelihood of the generated outputs being produced by the baseform for at least one other word.
-
2. A method of decoding a vocabulary word from outputs selected from an alphabet of outputs in response to a communicated word input, wherein each word in the vocabulary is represented by at least one probabilistic model, each probabilistic model having (i) stored transition probability values each representing the probability of a corresponding transition in a model being taken and (ii) stored output probability values each representing the probability of a corresponding output probability being produced at a given transition or transitions in a model, the method comprising the steps of:
-
(a) generating outputs in response to the communication of a known word input; and (b) biassing at least some of the stored values to enhance the likelihood that the generated outputs are produced by the baseform for the known word relative to the respective likelihood of the generated outputs being produced by the baseform for at least one other word. - View Dependent Claims (3)
-
-
4. In a speech recognition system in which labels, from an alphabet of labels, are generated by an acoustic processor at successive label times in response to a speech input and in which words or portions thereof are represented probabilistically by Markov models, wherein each Markov model is characterized by (i) states, (ii) transitions between states, and (iii) probabiltity items wherein some probability items have previously defined probability values θ
- '"'"' which correspond to the likelihood of a transition in a given model being taken and wherein other probability items have previously defined probability values θ
'"'"' which correspond to the likelihood of a specific label being produced at a transition of one or more predefined transitions in a given model, a method of evaluating counts from which enhanced probability values are derived comprising the steps of;(a) storing for each probability item a preliminary value θ
'"'"';(b) defining and storing a set of counts wherein each probability item is determined from the value of at least one count associated therewith in storage, each count in the set having a value corresponding to the probability of a specific transition τ
i being taken from a specific state Sj given (i) a specific label interval time t, (ii) a specific string of generated labels, and (iii) the stored θ
'"'"' values;(c) uttering a known subject word and generating outputs in response thereto; (d) selecting an incorrect word other than the known word and, for each count used in deriving the value of a probability item in said incorrect word model, determining a minus count value from the generated outputs of the uttered known word; and (e) defining an adjusted count value wherein the stored value of each count serves as an addend and the minus value of each count serves as a subtrahend thereof. - View Dependent Claims (5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20)
- '"'"' which correspond to the likelihood of a transition in a given model being taken and wherein other probability items have previously defined probability values θ
-
21. In a speech recognition system which decodes a vocabulary word from a string of output labels, each output label being selected from an alphabet of output labels in response to an uttered word input wherein each word in the vocabulary is represented by a baseform of at least one probabilistic finite state machine and wherein each probabilistic machine has transition probability items and output probability items, apparatus for determining probability values for probability items comprising:
-
means for storing a current probability value for each probability item; and means for biassing the stored current probability values to enhance the likelihood that outputs generated in response to the utterance of a known spoken word input are produced by the baseform for the known word relative to the respective likelihood of the generated outputs being produced by the baseform for at least one other word.
-
Specification