METHOD FOR SPEECH ROCOGNITION

• US 20100286984A1
• Filed: 06/18/2008
• Published: 11/11/2010
• Est. Priority Date: 07/18/2007
• Status: Active Grant

• Current Assignee: Cerence Operating Company (Cerence Inc.)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Perez, Jesus Fernando Guitarte, Wandinger, Michael, Littel, Bernhard
• Granted Patent: US 8,527,271 B2
• Time in Patent Office: Days
• US Class Current: 704/251
• CPC Class Codes: G01C 21/3608 using speech input, e.g. using speech recognition
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 2015/223 Execution procedure of a spoken command

Abstract

A method for the voice recognition of a spoken expression to be recognized, comprising a plurality of expression parts that are to be recognized. Partial voice recognition takes place on a first selected expression part, and depending on a selection of hits for the first expression part detected by the partial voice recognition, voice recognition on the first and further expression parts is executed.

Claims

1-22. -22. (canceled)

23. A method for speech recognition of a spoken expression, which exhibits first and remaining expression parts to be recognized, comprising:

conducting a partial speech recognition on the first expression part using a Hidden-Markov-Model, wherein the remaining expression parts are covered by fill data, the partial speech recognition using a first vocabulary having a list of search terms for the first expression part, the partial speech recognition selecting N best fitting hits for the first expression part from the list of search terms; and

conducting a complete speech recognition after the partial speech recognition, the complete speech recognition being performed for the first and the remaining expression parts, the complete speech recognition being conducted using a second vocabulary comprising a list of search term combinations depending on the N best fitting hits for the first expression part.

24. The method according to claim 23, wherein the first expression part is selected from the spoken expression based on a temporal energy contour of the spoken expression.

25. The method according to claim 23,

wherein the first expression part is selected from the spoken expression based on a speech activity recognition in the expression.

26. The method according to claim 23,

wherein

the spoken expression is recorded,

the first vocabulary comprises a first list of search terms, phoneme combinations and/or transition possibilities,

the second vocabulary comprises a list of search term combinations, phoneme combinations and/or transition probabilities, and

the complete speech recognition is conducted on the recorded expression to determine a selection of hits from the list of search term combinations.

27. The method according to claim 26, wherein the best fitting hits are the search terms from the first vocabulary which correspond most closely to the first expression part.

28. The method according to claim 26, wherein the selection of hits is the search term combinations of the second vocabulary which correspond most closely to the recorded expression.

29. The method according to claim 26, wherein the second vocabulary comprises search term combinations which are created by assigning additional terms and/or additional term sequences to the best fitting hits from the partial speech recognition.

30. The method according to claim 23, wherein the method recognizes the spoken expression using a hierarchically built database in which a generic term has further terms assigned thereto, and the further terms have even further subtopics are assigned thereto.

31. The method according to claim 30, wherein

the generic term comprises a city name,

the further terms comprise street names, and

the subtopics comprise house numbers.

32. The method according to claim 26, wherein

the second vocabulary comprises a subset of a complete vocabulary for the spoken expression, wherein the subset includes at least some of the best fitting hits with their corresponding further search terms of the complete vocabulary and excludes the rest of the complete vocabulary for the spoken expression.

33. The method according to claim 26, wherein after conducting the partial speech recognition, the first vocabulary is deleted from a memory and the second vocabulary is loaded into the memory.

34. The method according to claim 23, wherein

characteristics of the expression and/or of the first expression part are determined during partial speech recognition,

the characteristics determined comprise at least one of Mel Frequency Cepstral Coefficients (MFCC)-features and calculated emission probabilities, and

the characteristics determined during partial speech recognition are reused in complete speech recognition.

35. The method according to claim 23, wherein the first expression part is located at the beginning of the spoken expression.

36. The method according to claim 23 wherein at least parts of said second vocabulary are precompiled.

37. The method according to claim 32 wherein

a partial vocabulary is precompiled for the second vocabulary, and

reference vectors of a Hidden-Markov-Model are calculated and stored for the partial vocabulary.

38. A speech recognition device to recognize a spoken expression, which exhibits first and remaining expression parts to be recognized, the speech recognition device comprising:

a database including a memory and providing search terms and/or a phonemic transcription of the search terms; and

a control unit connected to the database to;

conduct a partial speech recognition on the first expression part using a Hidden-Markov-Model, wherein the remaining expression parts are covered by fill data, the partial speech recognition using a first vocabulary having a list of search terms for the first expression part, the partial speech recognition selecting N best fitting hits for the first expression part from the list of search terms; and

conduct a complete speech recognition after the partial speech recognition, the complete speech recognition being performed for the first and the remaining expression parts, the complete speech recognition being conducted using a second vocabulary comprising a list of search term combinations depending on the N best fitting hits.

39. The speech recognition device according to claim 38, wherein the speech recognition device is embedded in a motor vehicle.

40. The speech recognition device according to claim 39, wherein said speech recognition device is part of a vehicle navigation system.

41. The speech recognition device according to claim 39, wherein said speech recognition device is formed as a part of a mobile phone and/or of an audio player.

42. A motor vehicle, comprising a speech recognition device to recognize a spoken expression, which exhibits first and remaining expression parts to be recognized, the speech recognition device comprising:

a database including a memory and providing search terms and/or a phonemic transcription of the search terms; and

a control unit connected to the database to;

conduct a partial speech recognition on the first expression part using a Hidden-Markov-Model, wherein the remaining expression parts are covered by fill data, the partial speech recognition using a first vocabulary having a list of search terms for the first expression part, the partial speech recognition selecting N best fitting hits for the first expression part from the list of search terms; and

conduct a complete speech recognition after the partial speech recognition, the complete speech recognition being performed for the first and the remaining expression parts, the complete speech recognition being conducted using a second vocabulary comprising a list of search term combinations depending on the N best fitting hits.

43. A computer readable storage medium storing a program, which when executed by a computer performs a method for speech recognition of a spoken expression, which exhibits first and remaining expression parts to be recognized, the method comprising:

conducting a partial speech recognition on the first expression part using a Hidden-Markov-Model wherein the remaining expression parts are covered by fill data, the partial speech recognition using a first vocabulary having a list of search terms for the first expression part, the partial speech recognition selecting N best fitting hits for the first expression part from the list of search terms; and

conducting a complete speech recognition after the partial speech recognition, the complete speech recognition being performed for the first and the remaining expression parts, the complete speech recognition being conducted using a second vocabulary comprising a list of search term combinations depending on the N best fitting hits.