Sensor based approach recognizer selection, adaptation and combination

1. A method for operating a speech recognition system, in which a program-controlled recognizer performs the steps of:

dissecting a speech signal into frames and computing any kind of feature vector for each frame;

labeling frames by characters or groups of them yielding a plurality of labels per phoneme; and

decoding said labels according a predetermined acoustic model to construct one or more words or fragments of a word,

wherein a plurality of recognizers are accessible to be activated for speech recognition, and are combined in order to balance the results of speech recognition done by a single recognizer, the method further comprising:

a) collecting selection base data characterizing speech recognition boundary conditions with sensor means;

b) using program-controlled arbiter means for evaluating the collected data; and

c) selecting the best suited recognizer or a combination thereof out of the plurality of available recognizers according to said evaluation.

2. The method according to claim 1 , in which said sensor means is one or more of:

a decision logic, including software program, physical sensors or a combination thereof.

3. The method according to claim 1 , further comprising the steps of:

a) processing a physical sensor output in a decision logic implementing one or more of statistical tests, decision trees, and fuzzy membership functions; and

b) returning from said process a confidence value to be used in the sensor select/combine decision.

4. The method according to claim 1 , in which selection base data which have led to a recognizer select decision, is stored in a database for a repeated fast access thereof in order to obtain a fast selection of recognizers.

5. The method according to claim 1 , further comprising the step of:

selecting the number and/or combination of recognizers dependent of the current processor load.

6. The method according to claim 1 , further comprising the step of:

storing the mapping rule how one acoustic model is transformed to another one, instead of storing a plurality of models themselves.

7. A computer system comprising:

means for dissecting a speech signal into frames and computing any kind of feature vector for each frame;

means for labeling frames by characters or groups of them yielding a plurality of labels per phoneme;

means for decoding said labels according a predetermined acoustic model to construct one or more words or fragments of a word,

wherein a plurality of recognizers are accessible to be activated for speech recognition, and are combined in order to balance the results of speech recognition done by a single recognizer, the computer system carrying out the method of:

a) collecting selection base data characterizing speech recognition boundary conditions with sensor means;

b) using program-controlled arbiter means for evaluating the collected data; and

c) selecting the best suited recognizer or a combination thereof out of the plurality of available recognizers according to said evaluation.

8. The system according to claim 7 , in which said sensor means is one or more of:

a decision logic, a software program, physical sensors or a combination thereof.

9. A computer program for execution in a data processing system comprising computer program code portions for performing respective steps:

dissecting a speech signal into frames and computing any kind of feature vector for each frame;

labeling frames by characters or groups of them yielding a plurality of labels per phoneme; and

decoding said labels according a predetermined acoustic model to construct one or more words or fragments of a word,

wherein a plurality of recognizers are accessible to be activated for speech recognition, and are combined in order to balance the results of speech recognition done by a single recognizer, the method further comprising:

a) collecting selection base data characterizing speech recognition boundary conditions with sensor means;

b) using program-controlled arbiter means for evaluating the collected data; and

c) selecting the best suited recognizer or a combination thereof out of the plurality of available recognizers according to said evaluation,

when said computer program code portions are executed on a computer.

10. The computer program according to claim 9 , in which said sensor means is one or more of:

a decision logic, a software program, physical sensors or a combination thereof.

11. A computer program product stored on a computer usable medium comprising computer readable program means for causing a computer to perform the steps of:

dissecting a speech signal into frames and computing any kind of feature vector for each frame;

labeling frames by characters or groups of them yielding a plurality of labels per phoneme; and

decoding said labels according a predetermined acoustic model to construct one or more words or fragments of a word,

wherein a plurality of recognizers are accessible to be activated for speech recognition, and are combined in order to balance the results of speech recognition done by a single recognizer, the method further comprising:

a) collecting selection base data characterizing speech recognition boundary conditions with sensor means;

b) using program-controlled arbiter means for evaluating the collected data; and

c) selecting the best suited recognizer or a combination thereof out of the plurality of available recognizers according to said evaluation,

when said computer program product is executed on a computer.

12. The computer program product according to claim 11 , in which said sensor means is one or more of:

a decision logic, a software program, physical sensors or a combination thereof.