Speech recognition system

A method of performing an acoustic match of words in a vocabulary against a string of labels which represent a speech input, the method comprising the steps of:

• entering as inputs to the phone machine of a given phone (a) the string of labels and (b) a respective start-time distribution for the given phone;
• and
• generating (a) an end-time distribution and (b) a match value of the given phone relative to the entered labels based on the generated end-time distribution;
• the match value generated for each particular phone corresponding to the probability that said each particular phone produced the entered string of labels.

Apparatus for matching words with a string of incoming labels in a pattern recognition system, the apparatus comprising:

• at least one phone machine;
• each phone machine being characterized by having (a) a plurality of states and transition paths between states, (b) transition probabilities T(i.fwdarw.j) representing the probability of state S.sub.j given a current state S.sub.i where S.sub.i and S.sub.j may be the same state or different states, and (c) actual label probabilities wherein each actual label probability p(y.sub.k) indicates the probability that a label y.sub.k is produced by a given phone machine at a given transition from one state to a subsequent state where k is a label identifying notation;
• and
• each phone machine including (a) means for assigning to each y.sub.k in said each phone machine a single value p'(y.sub.k) and (b) means for replacing each actual output probability p(y.sub.k) at each transition in a given phone machine by the single value p'(y.sub.k) assigned to the corresponding y.sub.k.

A method of determining at least one word in a vocabulary having the highest probability of having generated a given string of incoming labels that were produced in response to a speech input, the method comprising the steps of:

• characterizing each word as a sequence of phonetic elements, wherein each phonetic element has (a) a start-time distribution of probabilities q.sub.n corresponding to respective successive start times t.sub.n, (b) a plurality of states between which transitions occur, (c) a plurality of transition probabilities, each indicating the probability that a given transition in a given phonetic element occurs, (d) a plurality of actual label probabilities, each actual output probability indicating the probability that a particular phonetic element generates a particular label at a particular transition in the particular phonetic element;
• and
• forming an approximate match for a subject word including the steps of:
• replacing all actual probabilities associated with a given label generated by a given phonetic element at any transition therein with a corresponding specific replacement value;
• and
• determining for one phonetic element after another in the subject word the probability .PHI..sub.n of a phonetic element ending at a respective one of a plurality of successive end times t.sub.n as a function of start-time distribution, the probability of the phonetic element generating a label string of each of various lengths, and the replacement value p'(y.sub.k) for each respective label y.sub.k that is to be generated by the phonetic element to produce the incoming string of labels.

A method of performing an approximate acoustic match between incoming labels and words in a vocabulary where each word is represented by a sequence of Markov model phone machines, the Markov model of each phone machine including states and transitions therebetween, the method comprising the steps of:

• determining the probability of being at a subject state at a given time, which includes the steps of (a) identifying each previous state that has a transition which leads to the subject state and determining the respective probability of each such previous state;
• (b) recognizing, for each such previous state, a value representing the probability of the label that must be generated at the transition between each such previous state and the current state in order to conform to the label string;
• and (c) combining the probability of each previous state and the respective value representing the label probability to provide a subject state probability over a corresponding transition;
• wherein, for each label in a given phone, setting the probability thereof to be a specific value throughout the given phone;
• determining the overall probability of being at the subject state from the subject state probabilities over all transitions leading thereto;
• and
• combining the overall probabilities for a given phone in a word to determine the likelihood of the given phone producing the string of labels.

In a speech recognition system, a method of measuring the likelihood of a word corresponding to a spoken input where the word is from a vocabulary of words, the method comprising the steps of:

• (a) generating a string of labels in response to a spoken input, each label (i) being from an alphabet of labels and (ii) representing a respective sound type;
• (b) determining label votes, each label vote representing the likelihood that a respective label is produced when a given word is uttered;
• and
• (c) for a subject word, accumulating a label vote for each of at least some of the labels generated in the string;
• the accumulated label votes providing information indicative of the likelihood of the subject word.

A method of evaluating the likelihood of a word corresponding to a speech input in a speech-recognition system comprising the steps of:

• for a subject word in a vocabulary of words, generating a first word score representing the subject word likelihood based on an acoustic match first algorithm;
• for the subject word, generating a second word score based on a second independent algorithm which differs from the first algorithm;
• and
• forming a total word score for the subject word from at least the first word score and the second word score.

In a speech recognition system in which each word is represented as a sequence of phones and in which an acoustic processor generates a string of successive labels in response to the utterance of speech, wherein each label corresponds to one of an alphabet of pre-defined sound types, a machine-implemetable method of determining a vote of each label of the alphabet for each vocabulary word, the method comprising the steps of:

• generating, in the acoustic processor, a string of labels in response to the uttering of a known script of sequential phones;
• evaluating a count indicative of the number of times each label in the label alphabet is generated for a given phone in response to the utterance of the known script;
• and
• repeating the evaluating step for each label as applied to each phone uttered in the known script.