System and method for unsupervised discovery of similar audio events
View Patent ↗A system and method are presented for unsupervised discovery of similar audio events collected from an automated dialing campaign. Locality Sensitive Hashing (LSH) is used to search for similar audio clips within a large dataset of audio recordings. A database is queried for possible matches between an unknown audio clip and any reference carrier audio message present in the database. The database is updated when new, or changed, carrier audio messages are detected.
1. A method for identifying new carrier audio messages, the method comprising the steps of:
determining, via a first process, a presence of one or more groups of similar audio recordings within a dataset of audio recordings, wherein each of the one or more groups of similar audio recordings comprises at least a threshold number of the audio recordings having a predetermined level of like characteristic; and
labeling each of the one or more groups of similar audio recordings as a one of the new carrier audio messages;
wherein the first process comprises the steps of:
for each of the audio recordings in the dataset of audio recordings, separating an associated audio signal into a signaling part and a speech part;
segmenting the speech part into frames;
computing feature vectors for each of the frames;
grouping the feature vectors into overlapping blocks, wherein:
each of the overlapping blocks comprises a plurality of the frames; and
each of the plurality of frames comprises a plurality of feature vectors;
using an algorithm to search the dataset of audio recordings for each of the overlapping blocks, wherein the search comprises:
encoding a first one of the overlapping blocks into a binary code,
searching the dataset of audio recordings for a hash bucket associated with the binary code, wherein if there is a hash bucket associated with the binary code, identifying the associated hash bucket, and
repeating the search until the search is performed for each of the overlapping blocks.
2. The method of claim 1 , wherein the signaling part comprises a ring tone and the speech part meets a predetermined threshold of length.
3. The method of claim 1 , wherein the feature vectors are invariant to signal degradations.
4. The method of claim 3 , wherein the feature vectors comprise Mel Frequency Cepstral Coefficients.
5. The method of claim 1 , wherein each of the overlapping blocks contains 10 of the frames, and the 10 of the frames each contain 13 of the feature vectors.
6. The method of claim 1 , wherein the each of the one or more groups of similar audio recordings comprises at least 5 of the audio recordings having the predetermined level of like characteristics.
7. The method of claim 1 , wherein the algorithm is a locality sensitive hashing algorithm.
8. The method of claim 1 , wherein the method further comprises the step of adding the binary code to an existing hash bucket in a database.
9. The method of claim 1 , wherein the method further comprises the step of creating a new hash bucket if an existing hash bucket associated with the binary code is not found.
10. A system for identifying new carrier audio messages, the system comprising:
a processor; and
a memory in communication with the processor, the memory storing instructions that, when executed by the processor, causes the processor to determine a presence of one or more groups of similar audio recordings with a dataset of audio recordings, wherein each of the one or more groups of similar audio recordings comprises at least a threshold number of the audio recordings having a predetermined level of like characteristic, and labeling each of the one or more groups of similar audio recordings as a one of the new carrier audio messages by:
for each of the audio recordings in the dataset of audio recordings, separating an associated audio signal into a signaling part and a speech part;
segmenting the speech part into frames;
computing feature vectors for each of the frames;
grouping the feature vectors into overlapping blocks, wherein:
each of the overlapping blocks comprises a plurality of the frames; and
each of the plurality of frames comprises a plurality of feature vectors;
using an algorithm to search the dataset of audio recordings for each of the overlapping blocks, wherein the search comprises:
encoding a first one of the overlapping blocks into a binary code,
searching the dataset of audio recordings for a hash bucket associated with the binary code, wherein if there is a hash bucket associated with the binary code, identifying the associated hash bucket, and
repeating the search until the search is performed for each of the overlapping blocks.
11. The system of claim 10 , wherein the signaling part comprises a ring tone and the speech part meets a predetermined threshold of length.
12. The system of claim 10 , wherein the feature vectors are invariant to signal degradations.
13. The system of claim 12 , wherein the feature vectors comprise Mel Frequency Cepstral Coefficients.
14. The system of claim 10 , wherein each of the overlapping blocks contains 10 of the frames, and the 10 of the frames each contain 13 of the feature vectors.
15. The system of claim 10 , wherein the each of the one or more groups of similar audio recordings comprises at least 5 of the audio recordings having the predetermined level of like characteristics.
16. The system of claim 10 , wherein the algorithm is a locality sensitive hashing algorithm.
17. The system of claim 10 , wherein the process further comprises adding the binary code to an existing hash bucket in a database.
18. The system of claim 10 , wherein the process further comprises creating a new hash bucket if an existing hash bucket associated with the binary code is not found.