Method and apparatus for training speech recognition algorithms for directory assistance applications

• US 5,488,652 A
• Filed: 04/14/1994
• Issued: 01/30/1996
• Est. Priority Date: 04/14/1994
• Status: Expired due to Term

• Current Assignee: Volt Delta Resources LLC (NewNet Communication Technologies LLC)
• Original Assignee: Northern Telecom Limited (Nortel Networks Corporation)
• Inventors: Sharp, R. Douglas, Bielby, Gregory J., Lennig, Matthew, Gupta, Vishwa N., Wasmeier, Hans A., Hodgson, Lauren C.
• Primary Examiner(s): Hofsass, Jeffery A.
• Assistant Examiner(s): WEAVER, SCOTT LOUIS
• Application Number:
  US08/227,830
• Time in Patent Office: 656 Days
• Field of Search: 379/67, 379/88, 379/89, 379/211, 379/213, 379/212, 379/201, 379/115, 379/142, 379/70, 379/84, 381/42, 381/43, 395/2, 395/2.51, 395/2.53, 395/2.65, 395/2.79, 395/2.84, 395/2.4, 395/2.43, 395/2.60
• US Class Current: 379/88.03
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum
• CPC Class Codes: H04M 2201/40 using speech recognition speech recognition per se G10L15/00
• CPC Class Codes: H04M 2242/22 Automatic class or number identification arrangements
• CPC Class Codes: H04M 3/42093 Notifying the calling party of information on the called or connected party
• CPC Class Codes: H04M 3/42102 Making use of the called party identifier
• CPC Class Codes: H04M 3/4931 Directory assistance systems
• CPC Class Codes: H04M 3/51 Centralised call answering arrangements requiring operator intervention , e.g. call or contact centers for telemarketing
• CPC Class Codes: H04Q 3/72 Finding out and indicating number of calling subscriber

Abstract

In methods and apparatus for at least partially automating a telephone directory assistance function, directory assistance callers are prompted to speak locality or called entity names associated with desired directory listings. A speech recognition algorithm is applied to speech signals received in response to prompting to determine spoken locality or called entity names. Desired telephone numbers are released to callers, and released telephone numbers are used to confirm or correct at least some of the recognized locality or called entity names. Speech signal representations labelled with the confirmed or corrected names are used as labelled speech tokens to refine prior training of the speech recognition algorithm. The training refinement automatically adjusts for deficiencies in prior training of the speech recognition algorithm and to long term changes in the speech patterns of directory assistance callers served by a particular directory assistance installation. The methods can be generalized to other speech recognition applications.

Claims

1. A method for performing a telephone directory assistance function, comprising:

for each of a plurality of callers;

prompting the caller to speak a name associated with a desired directory listing;

determining a telephone number desired by the caller based on a speech signal received from the caller subsequent to the step of prompting the caller;

releasing the desired telephone number to the caller; and

creating a call record comprising the released telephone number and a representation of the speech signal;

processing according to a parameter modification algorithm a plurality of the call records to modify parameters of a speech recognition algorithm; and

employing the speech recognition algorithm to recognize names spoken by callers in directory assistance calls received subsequent to modification of the parameters of the speech recognition algorithm.

2. A method for performing a telephone directory assistance function, comprising:

for each of a plurality of callers;

prompting the caller to speak a name associated with a desired directory listing;

receiving a speech signal from the caller;

applying a speech recognition algorithm to the received speech signal to recognize a name in a name vocabulary as corresponding to the received speech signal;

determining a telephone number desired by the caller using the recognized name;

releasing the desired telephone number to the caller; and

creating a call record comprising the released telephone number and a stored representation of the speech signal to associate the stored representation of the speech signal with the released telephone number;

processing according to a parameter modification algorithm a plurality of the call records to modify parameters of the speech recognition algorithm, said processing comprising;

processing each of the call records by;

deriving a plurality of name candidates from the released telephone number;

for each name candidate, calculating a measure of probability that the name candidate corresponds to the stored representation of the speech signal;

identifying a best name candidate having the highest measure of probability of the plurality of name candidates;

applying acceptance criteria based on the calculated measures of probability and, in response to the acceptance criteria being met, creating a training record comprising the stored representation of the speech signal and the best candidate name; and

processing the training records according to a training algorithm to modify the parameters of the speech recognition algorithm by, for each name of the name vocabulary;

accumulating statistics for said name derived from plural training records containing said name as a best name candidate; and

recalculating parameters associated with said name by the speech recognition algorithm from the statistics accumulated for said name; and

employing the speech recognition algorithm to recognize names spoken by callers in directory assistance calls received subsequent to modification of the parameters of the speech recognition algorithm.

3. A method as defined in claim 2, wherein:

the step of prompting the caller comprises prompting the caller to speak the name of a locality corresponding to the desired listing; and

the step of deriving the plurality of name candidates comprises deriving a plurality of locality name candidates from the released telephone number.

4. A method as defined in claim 3, wherein the step of deriving the plurality of locality name candidates from the released telephone number comprises deriving the plurality of locality name candidates from an NPA-NXX portion of the released telephone number.

5. A method as defined in claim 2, wherein:

the step of prompting the caller to speak a name comprises prompting the caller to speak the name of a subscriber corresponding to the desired listing;

the step of deriving the plurality of name candidates from the telephone number comprises deriving a plurality of subscriber name candidates from the released telephone number.

6. A method as defined in claim 2, wherein the step of determining the telephone number desired by the caller comprises calculating, according to the speech recognition algorithm, measures of probability that the speech signal received from the caller in response to the prompting corresponds to each of a plurality of names in the name vocabulary.

7. A method as defined in claim 6, wherein the step of determining the telephone number desired by the caller further comprises selecting as a best candidate name a name having the highest calculated measure of probability of the names in the plurality of names in the name vocabulary.

8. A method as defined in claim 7, wherein the step of determining the telephone number desired by the caller further comprises recognizing the name selected as the best candidate name as corresponding to the stored representation of the speech signal only if the calculated measures of probability meet predefined acceptance criteria.

9. A method as defined in claim 8, wherein the step of determining the telephone number desired by the caller comprises using the recognized name as a parameter in a database search for the desired telephone number.

10. A method as defined in claim 6, wherein:

the step of creating the call record comprises storing a plurality of names selected from the name vocabulary, the selected names having the highest measures of probability as calculated during the step of determining the telephone number desired by the caller; and

the step of creating the training record further comprises calculating further measures of probability that the stored representation of the speech signal in the call record corresponds to each of the plurality of names stored in the call record.

11. A method as defined in claim 10, wherein:

the step of creating the call record further comprises storing a name associated with the released telephone number in the call record; and

the step of creating the training record further comprises calculating a measure of probability that the stored representation of the speech signal in the call record corresponds to the name associated with the released telephone number stored in the call record.

12. A method as defined in claim 2, wherein the step of creating the training record is completed only if the calculated measures of probability meet predefined acceptance criteria.

13. A method as defined in claim 2, wherein the step of processing the training records according to the training algorithm comprises calculating model parameters for hidden Markov models associated with names in the name vocabulary according to a Viterbi training algorithm which is responsive to the stored representations of the speech signals and the best candidate names of the training records.

14. A method as defined in claim 13, wherein the step of processing the training records according to the training algorithm further comprises calculating model parameters for probability models which estimate probabilities of correct recognition of names in the name vocabulary by the speech recognition algorithm as functions of acceptance criteria parameters derived from the stored representations of speech signals and the hidden Markov models.

15. A method as defined in claim 14, wherein the step of processing the training records according to the training algorithm further comprises:

calculating rates of incorrect recognition for a plurality of threshold candidates; and

selecting as a threshold for acceptance criteria of the speech recognition algorithm one of the threshold candidates, the selected threshold candidate having a rate of incorrect recognition below a predetermined maximum rate of incorrect recognition.

16. A method as defined in claim 15, further comprising:

calculating a rate of incorrect recognition for each name transcription in the name vocabulary; and

labelling as decoys name transcriptions having a rate of incorrect recognition greater than a predetermined maximum rate of incorrect recognition.

17. A method for performing a telephone directory assistance function, comprising:

for each of a plurality of calls, each call being made by a respective caller from a respective terminal having a respective telephone number;

receiving a signal from an originating switch serving the respective terminal, the signal designating the telephone number of the respective terminal;

prompting the caller to speak a name associated with a desired directory listing;

determining a telephone number desired by the caller based on a speech signal received from the caller subsequent to the step of prompting the caller;

releasing the desired telephone number to the caller; and

creating a call record comprising the released telephone number and the telephone number of the respective terminal;

processing according to a parameter modification algorithm a plurality of the call records to modify parameters of a speech recognition algorithm, said processing comprising calculating model parameters for a plurality of a-priori probability models, each a-priori probability model representing an estimated probability of a respective name in a name vocabulary being spoken during a directory assistance call as a function of the telephone number of the terminal from which the directory assistance call is placed; and

employing the speech recognition algorithm to recognize names spoken by callers in directory assistance calls received subsequent to modification of the parameters of the speech recognition algorithm.

18. A method as defined in claim 17, wherein each a-priori probability model represents an estimated probability of a respective name in the name vocabulary being spoken during a directory assistance call as a function of an NPA-NXX portion of the telephone number of the terminal from which the directory assistance call is placed.

19. A method as defined in claim 17, wherein:

the step of creating a call record further comprises determining a respective call placement time for the call and storing the respective call placement time in the call record; and

each a-priori probability model represents an estimated probability of a respective name in the name vocabulary being spoken during a directory assistance call as a function of the telephone number of the terminal from which the directory assistance call is placed and of the call placement time.

20. Apparatus for performing a telephone directory assistance function, comprising:

a first processor for processing directory assistance calls, the first processor being operable to prompt callers to speak names associated with desired directory listings, to store in respective call records representations of speech signals received from the callers in response to the prompting, to execute a speech recognition algorithm to associate each stored representation of a speech signal with a respective name in a name vocabulary, and to record in the call records released telephone numbers received from a directory assistance database to associate the stored representations of the speech signals with respective ones of the released telephone numbers; and

a second processor for processing the call records created by the first processor, the second processor being operable to process the call records to modify parameters of the speech recognition algorithm.

21. Apparatus as defined in claim 20, wherein:

the second processor is operable for each of a plurality of the call records;

to derive at least one corresponding name from the released telephone number stored in the call record by searching a database;

to calculate for each derived name a respective measure of probability that the derived name corresponds to the stored representation of the speech signal in the call record;

to identify a best name candidate, the best name candidate being the derived name having a highest calculated measure of probability among the measures of probability calculated for the derived names;

to apply acceptance criteria based on the calculated measures of probability and, in response to the acceptance criteria being met, to create a training record comprising the stored representation of the speech signal and the best name candidate; and

the second processor is operable to process a plurality of the training records according to a training algorithm to modify parameters of the speech recognition algorithm.

22. Apparatus as defined in claim 21, wherein the first processor is operable to send indicia corresponding to recognized names to a directory assistance database for automated searching of the directory assistance database.

23. Apparatus for performing a telephone directory assistance function, comprising:

a first processor for processing directory assistance calls, the first processor being operable for each of a plurality of directory assistance calls to prompt a respective caller to speak a name associated with a desired directory listing, to store in a respective call record a representation of a speech signal received from the caller in response to the prompting, to execute a speech recognition algorithm to calculate respective measures of probability that the speech signal received from the caller corresponds to each of a plurality of names in a name vocabulary, and to record in the call record a respective released telephone number received from a directory assistance database to associate the stored representation of the speech signal with the released telephone number; and

a second processor for processing the call records created by the first processor, the second processor being operable to process the call records to modify parameters of the speech recognition algorithm and to download to memory accessible by the first processor the parameters of the speech recognition algorithm which have been modified according to the training algorithm.

24. Apparatus as defined in claim 23, wherein:

the first processor is operable for each of the plurality of directory assistance calls to store a respective plurality of indicia in the respective call record, the indicia corresponding to names having highest calculated measures of probability among the measures of probability calculated for the plurality of names in the name vocabulary for that directory assistance call; and

the second processor is operable for each of the plurality of call records to calculate further measures of probability that the respective stored representation of the speech signal in the call record corresponds to each indicia of the respective plurality of indicia stored in the call record.

25. Apparatus for performing a telephone directory assistance function, comprising:

a first processor for processing directory assistance calls, the first processor being operable to prompt callers to speak names associated with desired directory listings, to store in call records representations of speech signals received from the callers in response to the prompting, to execute a speech recognition algorithm to associate each stored representation of a speech signal with a respective name in a name vocabulary, to receive released telephone numbers from the directory assistance database and to record in the call records the released telephone numbers to associate the stored representations of the speech signals with respective ones of the released telephone numbers; and

a second processor for processing the call records created by the first processor, the second processor being operable to process the call records to modify parameters of the speech recognition algorithm, said processing comprising calculating further measures of probability that the stored representation of the speech signal in each of the call records corresponds to the released name in that call record.

26. Apparatus for performing a telephone directory assistance function, comprising:

a first processor for processing directory assistance calls, the first processor being operable to prompt callers to speak names associated with desired directory listings, to store in call records representations of speech signals received from the callers in response to the prompting, to execute a speech recognition algorithm to associate each stored representation of a speech signal with a respective name in a name vocabulary, to receive released telephone numbers from the directory assistance database, to store a call placement time in each of the call records; and

to record in the call records the released telephone numbers to associate the stored representations of the speech signals with respective ones of the released telephone numbers; and

a second processor for processing the call records created by the first processor, the second processor being operable to process the call records to modify parameters of the speech recognition algorithm based on the call placement times stored in the call records.

27. Apparatus for performing a telephone directory assistance function, comprising:

a first processor for processing directory assistance calls, the first processor being operable to prompt callers to speak names associated with desired directory listings, to store in call records representations of speech signals received from the callers in response to the prompting, to execute a speech recognition algorithm to associate each stored representation of a speech signal with a respective name in a name vocabulary, to receive released telephone numbers from a directory assistance database, and to record in the call records the released telephone numbers to associate the stored representations of the speech signals with respective ones of the released telephone numbers;

a second processor for processing the call records created by the first processor, the second processor being operable to process the call records to modify parameters of the speech recognition algorithm;

a first memory for storing a first set of instructions for execution by the first processor, the call records and a first set of the parameters of the speech recognition algorithm, the call records being read-accessible and the first set of the parameters of the speech recognition algorithm being write-accessible by the second processor; and

a second memory for storing a second set of instructions for execution by the second processor, a database containing telephone numbers and corresponding names and a second set of the parameters of the speech recognition algorithm, said database being read-accessible by the second processor to derive name candidates from the released telephone numbers.

28. Apparatus as defined in claim 27, further comprising:

a switch interface for interfacing the first processor to a switch of a switched telephone network; and

a directory assistance database interface for interfacing the first processor to the directory assistance database.

29. Apparatus as defined in claim 28, further comprising:

the switch of a switched telephone network connected to the first processor via the switch interface;

the directory assistance database connected to the first processor via the directory assistance database interface;

an operator terminal controller connected to the first processor via the switch interface and the switch, and connected to the directory assistance database via the switch; and

at least one operator terminal connected to the operator terminal controller.

30. Apparatus as defined in claim 29, further comprising an audio response unit connected to the directory assistance database and to the switch for audibly releasing the released telephone numbers of directory listings to the callers.

31. Apparatus as defined in claim 27, comprising a plurality of further processors operably connected to the second processor, each further processor having an associated memory, an associated switch interface and an associated directory assistance database interface.