System and method for recognizing speech

1. A method for recognizing speech including a sequence of words, comprising:

generating a set of interpretations of the speech using an acoustic model and a language model;

determining, for each interpretation, a score representing correctness of an interpretation in representing the sequence of words to produce a set of scores;

determining a constraint for recognizing the speech subject to a word sequence constraint;

determining a constraint factor indicating a degree of the consistency with the word sequence constraint;

determining a constrained scoring function based on the constraint factor for updating the set of scores; and

updating the set of scores using the scoring function; and

selecting, according to the updated set of scores, a best interpretation from the set of interpretations as the recognized speech, wherein steps of the method are performed by a processor.

2. The method of claim 1 , wherein the word sequence constraint includes one or combinations of a number of words in the sequence of words, a presence or absence of a specific word or a sequence of words, a time of utterance of the specific word, an order of at least two specific words in the sequence of words, a connection or separation of the two specific words in the sequence of words, a topic of the speech input.

3. The method of claim 1 , wherein the determining the constraint comprises:

communicating a subset of the set of interpretations to a user;

receiving the word sequence constraint in response to the communicating;

determining a type of the constraint based on the word sequence constraint; and

determining the constraint based on the type.

4. The method of claim 3 , wherein the type is a language type and the determining the constraint comprises:

updating the language model based on the word sequence constraint.

5. The method of claim 4 , wherein the word sequence constraint is a topic of the speech.

6. The method of claim 3 , wherein the type is an acoustic type and the determining the constraint comprises:

updating the acoustic model based on the word sequence constraint.

7. The method of claim 6 , wherein the acoustic model includes an alignment between words in the speech and acoustic features of the acoustic model.

8. The method of claim 7 , the constraint includes that there is only one word within a particular time region.

9. The method of claim 3 , wherein the type is a context type and the determining the constraint comprises:

determining a scoring function testing presence or absence of a specific word in each interpretation.

10. The method of claim 9 , wherein the scoring function tests for presence of the specific word, further comprising:

determining a direction of the speech based on the language model; and

updating the scoring function with a test for presence of words preceding and following the specific word according to the direction of the speech.

11. The method of claim 1 , wherein the scoring function S′(W|X) is

S

′

⁡

(

W

❘

X

)

∝

max

R

∈

𝒮

⁢

p

⁡

(

W

)

⁢

p

⁡

(

R

❘

W

)

⁢

p

⁡

(

X

❘

W

,

R

)

⁢

f

⁡

(

X

,

W

,

R

)

,

wherein ∝ denotes proportionality, W−(w 1 ) i=1 N is a sequence of N words w 1 , w 2 , . . . , w N in the interpretation, and w 1 is a word hypothesized in a position i, X={x j } j=1 T , denotes acoustic features of the acoustic model, where x j is an jth vector of the acoustic features of the speech input and T is the number of acoustic feature vectors, a function p(.|..) is a probability, S is a set of possible alignments, R−{r i } i=1 N is a set of hypothesized time regions of the speech for each hypothesized word, such that r i is the position hypothesized for word w 1 and ƒ(X,W,R) is a constraint factor outputting a number indicating a degree of consistency with the constraint for one or more of an acoustic sequence, a word sequence, and an alignment.

12. The method of claim 11 , further comprising:

determining an indicator function using the word sequence constraint; and

determining the constraint factor as a linear function of the indicator function with weight parameters that determine a degree of constraint satisfaction.

13. The method of claim 1 , wherein the constraint includes metadata of the sequence of words.

14. The method of claim 1 , further comprising:

determining the interpretation with the largest score as the recognized speech.

15. A method for recognizing speech of a user, comprising:

recognizing the speech to generate a set of interpretations associated with a corresponding set of scores representing correctness of each interpretation in representing the speech; and

updating iteratively the set of scores subject to at least one constraint, such that, for each iteration, a score of each interpretation is increased if the interpretation is consistent with the constraint, and is decreased if the interpretation is inconsistent with the constraint; and

selecting, according to the updated set of scores, an interpretation from the set of interpretations as the recognized speech, wherein steps of the method are performed by a processor, wherein the updating comprises:

communicating a subset of the set of interpretations to a user;

receiving a word sequence constraint in response to the communicating;

determining a type of the constraint based on the word sequence constraint, wherein the type is a context type;

determining a scoring function testing presence or absence of a specific word in each interpretation;

determining a direction of the speech based on a language model;

updating the scoring function with a test for presence of words preceding and following the specific word according to the direction of the speech; and

determining the constraint based on the type.

16. A system for recognizing speech, comprising:

a processor implementing a speech recognition module and an error correction module, wherein

the speech recognition module generates a set of interpretations of the speech input using an acoustic model and a language model, determines, for each interpretation, a score representing correctness of an interpretation in representing the speech and selects, according to the score of each interpretation, a best interpretation from the set of interpretation as the recognized speech; and wherein

the error correction module determines a constraint for recognizing the speech, and updates the score of each interpretation based on a consistency of the interpretation with the constraint, wherein the constraint is determined by:

communicating a subset of the set of interpretations to a user;

receiving a word sequence constraint in response to the communicating;

determining a type of the constraint based on the word sequence constraint, wherein the type is a context type;

determining a scoring function testing presence or absence of a specific word in each interpretation;

determining a direction of the speech based on a language model;

updating the scoring function with a test for presence of words preceding and following the specific word according to the direction of the speech; and

determining the constraint based on the type.

17. The system of claim 16 , further comprising:

an audio interface for receiving the speech representing a sequence of words;

a controller for communicating at least a subset of the set of interpretations to the user and for receiving a word sequence constraint from the user, wherein the processor determines the constraint based on the word sequence constraint.

18. The system of claim 16 , wherein the system for recognizing the speech is embedded in an instrumental panel of a vehicle.