Wordspotting system
View Patent ↗An approach to improving the performance of a wordspotting system includes providing an interface for interactive improvement of a phonetic representation of a query based on an operator identifying true detections and false alarms in a data set.
1. A computer-implemented method for searching for audio events comprising:
accepting a specification of a search term;
determining a first representation of the search term;
identifying a first plurality of putative instances of the search term in audio data according to the first representation;
determining by a processor a second representation of the search term according to information received responsive to a presentation of the first plurality of instances, wherein determining the second representation of the search term includes selecting the second representation of the search term from a plurality of representations determined from the accepted specification of the search term; and
identifying a second plurality of putative instances of the search term in the audio data according to the second representation.
2. The method of claim 1 further comprising: providing information relating the first plurality and the second plurality of putative instances of the search term.
3. The method of claim 1 further comprising iteratively refining a representation of the search term by repeating determining of a further representation of the search term according to information received responsive to a presentation of putative instances determined according to a prior representation of the search term.
4. The method of claim 1 wherein the specification of the search term comprises a text representation of a phrase including one or more words.
5. The method of claim 3 wherein the text representation comprises an orthographic representation.
6. The method of claim 1 wherein determining the first representation includes determining said representation in terms of subword units.
7. The method of claim 6 wherein determining the first representation includes determining a phonetic representation of the search term.
8. The method of claim 6 wherein determining the first representation includes applying at least one of a dictionary lookup and a text-to-phoneme procedure.
9. The method of claim 1 wherein identifying the first plurality of putative instances includes performing speech recognition on the audio data.
10. The method of claim 9 wherein performing speech recognition includes performing word spotting according to the first representation of the search term.
11. The method of claim 9 wherein at least some processing steps of the speech recognition are performed prior to accepting the specification of the search term.
12. The method of claim 11 wherein the at least some processing steps include phoneme-based analysis of the audio data.
13. The method of claim 1 further comprising: presenting a display of the first plurality of putative instances.
14. The method of claim 13 wherein presenting the display includes displaying an ordered display according to scores for the putative instances.
15. The method of claim 13 further comprising: accepting an indication of accuracy for each of one or more of the putative instances.
16. The method of claim 15 wherein the indication of accuracy includes at least one of an indication that the putative instance is a true instance and an indication that the putative instance is a false detection of the search term.
17. The method of claim 13 wherein presenting the display includes presenting audio representations of selected putative instances.
18. The method of claim 13 wherein presenting the display includes presenting a graphical representation indicating characteristics of subword units in the first representation of the search term.
19. The method of claim 18 wherein the characteristics of subword units include match quality of the units to the audio data.
20. The method of claim 18 wherein the characteristics of subword units include temporal characteristics of the subword units.
21. The method of claim 1 wherein determining the second representation of the search term includes computing the second representation according to an indication of accuracy for each of one or more of the first plurality of putative instances.
22. The method of claim 21 wherein computing the second representation is further according to the accepted specification of the search term.
23. The method of claim 22 wherein computing the second representation includes selecting the second representation from a set of possible representations of the search term.
24. The method of claim 23 wherein selecting the second representation includes selecting a representation that best represents putative instances indicated as being true instances of the search term.
25. The method of claim 23 wherein selecting the second representation includes selecting a representation that provides a best performance measure according to the indications of accuracy of the putative instances.
26. The method of claim 25 wherein the selected representation optimizes a figure of Merit.
27. The method of claim 1 further comprising: presenting a display of the second plurality of putative instances.
28. The method of claim 27 wherein presenting the display of the second plurality of instances includes presenting an association between putative instances in the second plurality and the first plurality of putative instances.
29. The method of claim 28 wherein presenting the association includes presenting a change of rank of a putative instance between use of the first representation and the second representation of the search term.
30. The method of claim 1 , wherein determining the second representation of the search term includes:
presenting the first plurality of putative instances of the search term to a user; and
determining the second representation of the search term according to information received from the user responsive to the presentation of the first plurality of putative instances.
31. A computer-implemented method for determining a representation of a search term comprising:
accepting a specification of the search term;
determining a first representation of the search term;
identifying a first plurality of putative instances of the search term in audio data according to the first representation;
presenting a display of the first plurality of putative instances; and
determining by a processor a second representation of the search term according to information received responsive to the presentation of the display of the first plurality of instances, wherein determining the second representation of the search term includes selecting the second representation of the search term from a plurality of representations determined from the accepted specification of the search term.
32. The method of claim 31 further comprising: accepting an indication of accuracy for each of one or more of the putative instances.
33. The method of claim 32 wherein the indication of accuracy includes at least one of an indication that the putative instance is a true instance and an indication that the putative instance is a false detection of the search term.
34. Software stored on a computer-readable medium comprising instructions for causing a data processing system to:
accept a specification of a search term; determine a first representation of the search term;
identify a first plurality of putative instances of the search term in audio data according to the first representation;
determine a second representation of the search term according to information received responsive to a presentation of the first plurality of instances, wherein the second representation of the search term is selected from a plurality of representations determined from the accepted specification of the search term; and
identify a second plurality of putative instances of the search term in the audio data according to the second representation.
35. A system for wordspotting comprising:
an interface;
a processor;
a query optimizer; and
a word spotting engine; wherein
the interface is configured to accept a specification of a search term; and
the query optimizer is configured to determine a first representation of the search term;
the word spotting engine is configured to identify a first plurality of putative instances of the search term in audio data according to the first representation;
the interface is configured to present the first plurality of putative instances and receive information responsive to a presentation of the first plurality of instances; and
the query optimizer is further configured to cause the processor to determine a second representation of the search term according to the information received responsive to the presentation of the first plurality of instances, wherein the second representation of the search term is selected from a plurality of representations determined from the accepted specification of the search term.
36. A computer-implemented method for searching for audio events comprising:
accepting a specification of a search term;
performing a first search of audio data according to a first representation of the search term to identify a first plurality of putative instances of the search term;
determining by a processor a second representation of the search term according to an indication of accuracy for each of one or more of the first plurality of putative instances of the search term, wherein determining the second representation of the search term includes selecting the second representation of the search term from a plurality of representations determined from the accepted specification of the search term; and
performing a second search of audio data according to the second representation of the search term to identify a second plurality of putative instances of the search term.