Compressed phonetic representation
View Patent ↗An audio processing system makes use of a number of levels of compression or data reduction, thereby providing reduced storage requirements while maintaining a high accuracy of keyword detection in the original audio input.
1. A method for searching acoustic information comprising:
processing at an input section an acoustic input over a time interval, including forming a variable rate compressed representation of a degree of presence of a set of subword units as a function of a time in the interval;
passing the compressed representation to a search section; and
processing the compressed representation at the search section to form a search output representing a presence of a query in the acoustic input,
wherein processing the acoustic input further includes forming a fixed length representation of the degree of presence of the set of subword units for each of a sequence of times.
2. The method of claim 1 wherein the subword units comprise phonetic units.
3. The method of claim 1 wherein processing the acoustic input further includes forming a variable compression of the fixed length representation at times in the sequence of times to form the variable rate compressed representation.
4. The method of claim 1 wherein processing the acoustic input further includes for each of the sequence of times identifying a set of likely subword units, and encoding information associated with the set separately from information associated with the remaining subword units in the compressed representation.
5. The method of claim 1 wherein processing the acoustic input further includes forming a vector quantization of the degree of presence of the set of subword units.
6. The method of claim 1 wherein processing the acoustic input further includes determining an indicator of interest in regions of the acoustic input, and forming the variable rate representation includes applying a different degree of compression according to the indicator of interest.
7. The method of claim 6 wherein the indicator of interest comprises a binary indicator, and applying the different degree of compression comprises omitting regions of non interest from the compressed representation.
8. The method of claim 1 wherein processing the acoustic input further includes forming a compression map that encodes an association of time in the time interval of the input and locations in the variable rate compression.
9. The method of claim 1 wherein processing the acoustic input further includes performing an initial search for additional units defined in terms of the subword units, and wherein the compressed representation encodes a degree of presence of a set of the additional units as a function of a time.
10. The method of claim 9 wherein the additional units comprise context-dependent phoneme units.
11. The method of claim 1 wherein processing compressed representation at the search section includes determining a degree of presence of a set of additional units as a function of time.
12. The method of claim 11 wherein the additional units comprise context-dependent phoneme units.
13. The method of claim 12 wherein the additional units depend on the query.
14. The method of claim 11 wherein determining the degree of presence of the additional units comprises accessing a statistically trained model that relates the degree of presence of the set of subword units to degree of presence of each of a set of context-dependent subword units.
15. The method of claim 1 wherein passing the compressed representation to the search section includes storing the compressed representation in a storage device.
16. The method of claim 1 wherein passing the compressed representation to the search section includes passing the compressed representation over a communication link between the input section and the search section.
17. The method of claim 1 wherein processing the compressed representation at the search section includes decompressing the variable rate compressed representation, and searching for the presence of the query using the decompressed representation.
18. The method of claim 17 wherein processing the compressed representation at the search section further includes accessing a compression map that encodes an association of time in the time interval of the input locations in the variable rate compression to select portions of the compressed representation for processing.
19. The method of claim 17 wherein processing the compressed representation at the search section further includes accessing a compression map that encodes an association of time in the time interval of the input locations in the variable rate compression to determine times in the time interval of the input associated with detected locations of the query.
20. The method of claim 1 wherein processing the compressed representation at the search section includes determining a degree of presence of a set of the additional units defined in terms of the subword units based on a decompression of the compressed representation, and using the determined degrees in forming the search output.
21. An audio search system comprising:
an input section for processing an acoustic input over a time interval, including forming a variable rate compressed representation of a degree of presence of a set of subword units as a function of a time in the interval, the input section including a phonetic analysis module for processing an acoustic input and forming a fixed length representation of the degree of presence of the set of subword units for each of a sequence of times;
a search section for processing the compressed representation to form a search output representing a presence of a query in the acoustic input; and
a data transfer section for passing the compressed representation from the input section to the search section.
22. The audio search system of claim 21 wherein the input section comprises:
a variable rate compression module for forming the variable rate compressed representation from the fixed-length representation.
23. The audio search system of claim 21 wherein the search section comprises:
a decompression module for reconstructing a fixed-length representation of the degree of presence of the set of subword units for each of a sequence of times; and
a search module for processing the fixed-length representation and a query to determine data representing presence of the query in the acoustic input.