System and method for outputting spoken information in response to input speech signals

• US 5,754,736 A
• Filed: 09/08/1995
• Issued: 05/19/1998
• Est. Priority Date: 09/14/1994
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Aust, Harald
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/526,017
• Time in Patent Office: 984 Days
• Field of Search: 395/2.61, 395/2.62, 395/2.66, 395/2.52, 395/2.44, 395/2.56, 395/2.63
• US Class Current: 704/252
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/193 Formal grammars, e.g. finite state automata, context free grammars or word networks

Abstract

The invention aims to make an automatic information system suitable for the input of requests which are spoken as naturally as possible and which are not subject to restrictions about which the user must be informed in advance. To this end, the system and the method are subdivided into several, essentially independently operating segments with unambiguous interfaces. A first segment serves for the recognition of words and outputs a word graph. A second segment forms a concept graph from the word graph in that essentially the word graph is searched only for word sequences wherefrom values for a database enquiry can be derived; such word sequences, or also single words, are replaced by concepts in the concept graph. In a further segment the concepts in the concept graph are bridged and gaps between the concepts are filled. A speech model is taken into account for this operation. A stochastic attributed grammar is used to determine the concepts from the word graph and to determine the meaning of the word sequences in the concepts. Because irrelevant filler words in the spoken request are ignored early on, the computational expenditure is comparatively small despite the freely pronounced requests.

Claims

1. A system for outputting spoken information in response to input speech signals, comprising:

first means for comparing the speech signals with stored reference signals which correspond to words, and for outputting words determined by the comparisons, together with a score and with statements as regards starting point and end point for each word output, and for storing the words output with the statements in order to form a word graph from consecutive words,

second means for replacing predetermined word sequences in the word graph by predetermined concepts with a score which is derived from the sum of the scores of the words of the word sequence replaced, and also with the same starting and end points as the relevant replaced word sequence in order to form a concept graph and for determining values from the replaced word sequences for the individual concepts in the concept graph,

third means for determining a sequence of concepts from the concept graph on the basis of the scores, and

fourth means for converting the values in the concepts of the sequence into data for a database enquiry in order to control a memory which contains predetermined words or word sequences, and for outputting selected word sequences in the form of complete sentences.

2. A system as claimed in claim 1, in which the first means are arranged to recognize speech pauses in the speech signal, to output these pauses with a score, and to replace each sequence of a word and a directly subsequent speech pause in the word graph by the word obtaining a supplemented score equal to the combination of the scores of the word and the speech pause.

3. A system as claimed in claim 1, in which the second means are arranged to generate additionally, for each concept, a filler concept and also a filler concept between pairs of two concepts in the concept graph which do not directly adjoin one another, each filler concept being given a score which is derived from the sum of the scores of the words in the word graph which have been replaced by the filler concept, no value being assigned to any filler concept.

4. A system as claimed in claim 3, in which the second means are arranged to supplement the score of the concepts in the concept graph by a concept-dependent additional score and the filler concepts by an additional score which is dependent on the length of the relevant filler concept.

5. A system as claimed in claim 4, in which the second means are arranged to supplement the score of the concepts in the concept graph by a score which is dependent on the combination of each concept with each directly preceding concept or with the directly preceding start of the speech signal.

6. A system as claimed in claim 1, in which the second means are arranged to replace further word sequences in the word graph by predetermined dummy concepts wherefrom no values can be determined for conversion into data for a database enquiry, at least one score being assigned to the dummy concepts.

7. A method for outputting spoken information in response to input speech signals, comprising the following steps:

comparing the speech signals with stored reference signals which correspond to words, and outputting words determined by the comparisons, together with a score and with statements as regards starting point and end point for each word output, and storing the words output with the statements in order to form a word graph from consecutive words,

replacing predetermined word sequences in the word graph by predetermined concepts with a score which is derived from the sum of the scores of the words of the word sequence replaced, and also with the same starting and end points as the relevant replaced word sequence forming a concept graph and determining values from the replaced word sequences for the individual concepts in the concept graph,

determining a sequence of concepts from the concept graph on the basis of the scores, and

converting the values in the concepts of the sequence into data for a database enquiry in order to control a memory which contains predetermined words or word sequences, and outputting selected word sequences in the form of complete sentences.

8. A method as claimed in claim 7, in which speech pauses in the speech signal are recognized and output with a score, and each sequence of a word and a directly subsequent speech pause in the word graph is replaced by the word obtaining a supplemented score equal to the combination of the scores of the word and the speech pause.

9. A method as claimed in claim 7, in which additionally, for each concept, a filler concept is generated and also a filler concept between pairs of two concepts in the concept graph which do not directly adjoin one another, and each filler concept is given a score which is derived from the sum of the scores of the words in the word graph which have been replaced by the filler concept, and no value is assigned to any filler concept.

10. A method as claimed in claim 9, in which the score of the concepts in the concept graph are supplemented by a concept-dependent additional score and the filler concepts are supplemented by an additional score which is dependent on the length of the relevant filler concept.

11. A method as claimed in claim 10, in which the score of the concepts in the concept graph are supplemented by a score which is dependent on the combination of each concept with each directly preceding concept or with the directly preceding start of the speech signal.

12. A method as claimed in claim 7, in which further word sequences in the word graph are replaced by predetermined dummy concepts wherefrom no values can be determined for conversion into data for a database enquiry and to which at least one score is assigned.