Method for automatically punctuating a speech utterance in a continuous speech recognition system

Claim 2

In a speech recognition system that outputs a sequence of decoded words, a method for automatically punctuating the sequence of decoded words comprises the steps of: in a vocabulary that defines items in a language model, the items including words and punctuation marks, assigning at least one baseform to each of the punctuation marks, the at least one baseform corresponding to at least one of silence and a non-word noise; in the language model, defining conditional probabilities for each of the punctuation marks based upon at least one of at least one preceding word and at least one succeeding word; and automatically inserting a subject punctuation mark at a given point in the sequence of decoded words when an acoustic score and a language model score associated with the subject punctuation mark produce a higher combined likelihood than the acoustic score and the language model score associated with any other item in the vocabulary for the given point in the sequence of decoded words.

Claim 3

The method of claim 1, wherein the baseform corresponding to the silence is represented by at least one silence phone.

Claim 4

The method of claim 1, wherein the baseform corresponding to the non-word noise is represented by one of at least one consonant phone preceded and followed by the silence phone, at least one vowel phone preceded and followed by the silence phone, and at least one consonant phone and at least one vowel phone preceded and followed by the silence phone.

Claim 5

The method of claim 1, wherein the language model is an n-gram language model, and said method further comprises the step of: in a training phase of the n-gram language model, segmenting a corresponding text corpus into paragraphs to obtain word frequency counts that include each of the punctuation marks in a position other than a last position.

Claim 6

The method of claim 4, wherein each paragraph includes about 200 to 500 words.

Claim 7

The method of claim 4, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 8

The method of claim 1, wherein the language model is an n-gram language model, and said method further comprises the step of: in a training phase of the n-gram language model, segmenting a corresponding text corpus into paragraphs to obtain the conditional probabilities for the punctuation marks.

Claim 9

The method of claim 7, wherein each paragraph includes about 200 to 500 words.

Claim 10

The method of claim 7, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 11

The method of claim 1, wherein at least one punctuation mark is also assigned at least one baseform corresponding to a pronunciation of a word representing the punctuation mark.

Claim 12

In a speech recognition system which recognizes a spoken utterance and, in response, outputs a sequence of decoded words, a method for automatically punctuating the sequence of decoded words comprises the steps of: assigning at least one baseform to a subject punctuation mark in a vocabulary of words and punctuation marks, the at least one baseform including at least one of a silence phone, two consecutive silence phones, at least one consonant phone preceded and followed by the silence phone, at least one vowel phone preceded and followed by the silence phone, and at least one consonant phone and at least one vowel phone preceded and followed by the silence phone, the vocabulary defining items for which language model scores exist; defining at least one conditional probability for the subject punctuation mark in a language model based upon at least one of at least one preceding word and at least one succeeding word; and generating a word match score corresponding to the probability that the subject punctuation mark corresponds to acoustic data generated by said utterance; generating a language model score corresponding to the probability that the subject punctuation mark corresponds to the acoustic data generated by said utterance; and automatically inserting the subject punctuation mark at a given position in the sequence of decoded words when the combination of the word match score and the language model score produces the highest combined likelihood over any other word and punctuation mark in the vocabulary for the given position.

Claim 13

In a speech recognition system which recognizes a spoken utterance and, in response, outputs a sequence of decoded words, a method for automatically punctuating the sequence of decoded words comprises the steps of: assigning at least one baseform to a subject punctuation mark in a vocabulary of words and punctuation marks, the at least one baseform including at least one of a silence phone, two consecutive silence phones, at least one consonant phone preceded and followed by the silence phone, at least one vowel phone preceded and followed by the silence phone, and at least one consonant phone and at least one vowel phone preceded and followed by the silence phone, the vocabulary defining items for which language model scores exist; defining at least one conditional probability for the subject punctuation mark in a language model based upon at least one of at least one preceding word and at least one succeeding word; and generating a word match score corresponding to the probability that the subject punctuation mark corresponds to acoustic data generated by said utterance; generating a language model score corresponding to the probability that the subject punctuation mark corresponds to the acoustic data generated by said utterance; and automatically inserting the subject punctuation mark at a given position in the sequence of decoded words when the combination of the word match score and the language model score produces the highest combined likelihood over any other word and punctuation mark in the vocabulary for the given position.

Claim 14

The method of claim 11, wherein each word, silence and punctuation mark in the vocabulary is represented by at least one baseform.

Claim 15

The method of claim 11, wherein the language model is an n-gram language model, and said method further comprises the step of: in a training phase of the n-gram language model, segmenting a corresponding text corpus into paragraphs to obtain word frequency counts that include the subject punctuation mark in a position other than a last position.

Claim 16

The method of claim 13, wherein each paragraph includes about 200 to 500 words.

Claim 17

The method of claim 13, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 18

The method of claim 11, wherein the language model is an n-gram language model, and said method further comprises the step of: in a training phase of the n-gram language model, segmenting a corresponding text corpus into paragraphs to obtain the at least one conditional probability for the subject punctuation mark.

Claim 19

The method of claim 16, wherein each paragraph includes about 200 to 500 words.

Claim 20

The method of claim 16, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 21

In a speech recognition system which recognizes a spoken utterance consisting of an input sequence of spoken words and outputs a sequence of decoded words, a method for automatically punctuating the decoded words comprises the steps of: training an acoustic model; training an n-gram language model; building a vocabulary of items, the items including words and punctuation marks, said building step including assigning at least one baseform to each of the punctuation marks in the vocabulary, the at least one baseform corresponding to at least one of silence and a non-word noise, the vocabulary defining the items for which language model scores exist; defining conditional probabilities for the punctuation marks in the language model based upon at least one of at least one preceding word and at least one succeeding word; and generating a word match score corresponding to the probability that a subject punctuation mark in the vocabulary corresponds to acoustic data generated by said utterance; generating a language model score corresponding to the probability that the subject punctuation mark corresponds to acoustic data generated by said utterance; and automatically inserting the subject punctuation mark at a given point in the sequence of decoded words when the combination of the word match score and the language model score produces the highest combined likelihood over any other item in the vocabulary for the given point.

Claim 22

In a speech recognition system which recognizes a spoken utterance consisting of an input sequence of spoken words and outputs a sequence of decoded words, a method for automatically punctuating the decoded words comprises the steps of: training an acoustic model; training an n-gram language model; building a vocabulary of items, the items including words and punctuation marks, said building step including assigning at least one baseform to each of the punctuation marks in the vocabulary, the at least one baseform corresponding to at least one of silence and a non-word noise, the vocabulary defining the items for which language model scores exist; defining conditional probabilities for the punctuation marks in the language model based upon at least one of at least one preceding word and at least one succeeding word; and generating a word match score corresponding to the probability that a subject punctuation mark in the vocabulary corresponds to acoustic data generated by said utterance; generating a language model score corresponding to the probability that the subject punctuation mark corresponds to acoustic data generated by said utterance; and automatically inserting the subject punctuation mark at a given point in the sequence of decoded words when the combination of the word match score and the language model score produces the highest combined likelihood over any other item in the vocabulary for the given point.

Claim 23

The method of claim 19, wherein the baseform corresponding to the silence is represented by at least one silence phone.

Claim 24

The method of claim 19, wherein the baseform corresponding to the silence is represented by two consecutive silence phones.

Claim 25

The method of claim 19, wherein the baseform corresponding to the non-word noise is represented by one of at least one consonant phone preceded and followed by the silence phone, at least one vowel phone preceded and followed by the silence phone, and at least one consonant phone and at least one vowel phone preceded and followed by the silence phone.

Claim 26

The method of claim 19, wherein the step of training the n-gram language model further comprises the step of segmenting a training text corpus into a plurality of paragraphs to obtain word frequency counts that include the subject punctuation mark in a position other than a last position.

Claim 27

The method of claim 23, wherein each paragraph includes about 200 to 500 words.

Claim 28

The method of claim 23, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 29

The method of claim 19, wherein the step of training the n-gram language model further comprises the step of segmenting a corresponding text corpus into paragraphs to obtain the conditional probabilities for the punctuation marks.

Claim 30

The method of claim 26, wherein each paragraph includes about 200 to 500 words.

Claim 31

The method of claim 26, wherein the n-gram language model is one of a bigram language model and a trigram language model.

Claim 32

The method of claim 19, wherein at least one punctuation mark is also assigned at least one baseform corresponding to a pronunciation of a word representing the punctuation mark.