Word-level blind diarization of recorded calls with arbitrary number of speakers

1. A method of creating an acoustic signature, the method comprising:

receiving audio data at a communication interface of a computing system on a frame by frame basis, wherein at least one frame includes metadata regarding the audio data, wherein the metadata identifies a first speaker and a second speaker associated with the audio data;

selecting a linguistic model to create a speech to text transcription according to the metadata;

creating a speech to text transcription of the audio data;

segmenting the audio data according to identified word sequences;

clustering segments of the audio data according to the identified word sequences;

applying an acoustical matching technique and a text analysis technique to each of the clustered segments to identify which clustered segments include audio that was likely spoken by the first speaker and which clustered segments include audio that was likely spoken by the second speaker; and

tagging the clustered segments to indicate whether they include audio spoken by the first speaker or the second speaker; and

creating an acoustic signature for the first speaker using some or all of the tagged clustered segments, wherein creating the acoustical signature for the first speaker comprises:

classifying some or all of the tagged clustered segments to identify a set of common speaker Gaussian mixture models (GMMs);

constructing a first super-GMM for the set of common speaker GMMs; and

constructing a second super-GMM for a set of generic speaker GMMs, wherein the acoustic signature comprises the first super-GMM and the second super-GMM.

2. The method according to claim 1 , further comprising using the metadata to select the linguistic model to create the speech to text transcription.

3. The method according to claim 1 , further comprising filtering out non-speech frames by evaluating envelope energy level of respective frames and comparing the envelope energy to a threshold energy above which a frame includes an utterance.

4. The method according to claim 1 , further comprising assigning start time frames to the word sequences, wherein the clustering is further based on respective time lengths of the word sequences.

5. The method according to claim 1 , wherein the segmenting further comprises using voice activity detection to segment the audio data into utterances having a statistical likelihood of emanating from a single speaker.

6. The method according to claim 1 , wherein the metadata comprises an identification number for the first speaker and the second speaker.

7. The method according to claim 1 , wherein the metadata comprises context data selected from the group consisting of including a topic, time, date, and location of audio data origin.

8. A method for creating a plurality of acoustic signatures and for performing diarization, comprising:

receiving audio data at a communication interface of a computing system, wherein at least one frame includes metadata regarding the audio data;

selecting a model to create a speech to text transcription according to the metadata;

creating the speech to text transcription of the audio data in accordance with the metadata;

clustering respective segments of the audio data according to word sequences;

classifying the segments to identify a set of common speaker Gaussian mixture models (GMMs) and a set of generic speaker GMMs, wherein the classifying includes constructing an undirected similarity graph having vertices corresponding to a plurality of speaker models of previously recorded audio sessions in a training set;

wherein the classifying further includes determining with a processor in a computing system a degree of similarity between the corresponding vertices in the undirected similarity graph in relation to at least one threshold degree of similarity;

generating an acoustic signature by at least:

constructing a super-GMM for the set of common speaker GMMs, and

constructing a super-GMM for the set of generic speaker GMMs by generating a set of random vectors and training a GMM over these random vectors, wherein the acoustic signature for respective common speakers is given as a super-model pair of the two constructed super-GMMS; and

storing the super-GMMs in a computing system memory;

receiving additional audio data at the communication interface;

identifying a respective common speaker using the super-model pair; and

labeling the additional audio data with an identified common speaker label.

9. The method according to claim 8 , further comprising utilizing a diagonal Gaussian distribution for the clusters to calculate a log likelihood that respective segments are within the cluster.

10. The method according to claim 8 , wherein the prerecorded training sets reside on a server.

11. The method according to claim 8 , further comprising filtering short utterances as background audio.

12. The method according to claim 8 , further comprising filtering out short utterances on a time duration basis.

13. The method according to claim 8 , further comprising classifying the clusters according to Mel-frequency cepstral coefficients (MFCC) for respective frames of audio data.

14. The method according to claim 8 , further comprising:

determining a cluster of segments to be comprised of respective utterances and representing a distribution of feature vectors in the respective utterances;

characterizing each feature vector in terms of its probability of being present in one of the clusters;

calculating a distance metric between utterances according to the probability;

identifying time between speakers in the audio stream.

15. The method according to claim 14 , further comprising:

using distances between utterances to construct an affinity matrix based upon respective distances;

computing a stochastic matrix from the affinity matrix;

computing eigenvalues and corresponding eigenvectors of the stochastic matrix; and

computing the embedding of the utterances into dimensional vectors;

identifying embedded utterances in a frame as an additional speaker or as additional background audio.

16. The method according to claim 8 , further comprising using the processor to determine the degree of similarity by calculating a distance (δ) between the corresponding vertices of the speaker models.

17. A system of creating an acoustic signature, the system comprising:

at least one computer connected for communication over a network and comprising a processor in data communication with memory storing computer readable commands configured to implement the following steps in a computerized process:

receiving audio data at a communication interface of the computer on a frame by frame basis, wherein at least one frame includes metadata regarding the audio data, wherein the metadata identifies a first speaker and a second speaker associated with the audio data;

selecting a linguistic model to create a speech to text transcription according to the metadata;

creating a speech to text transcription of the audio data;

segmenting the audio data according to identified word sequences;

clustering segments of the audio data according to the identified word sequences;

applying an acoustical matching technique and a text analysis technique to each of the clustered segments to identify which clustered segments include audio that was likely spoken by the first speaker and which clustered segments include audio that was likely spoken by the second speaker;

tagging the clustered segments to indicate whether they include audio spoken by the first speaker or the second speaker; and

creating an acoustic signature for the first speaker using some or all of the tagged clustered segments, wherein creating the acoustical signature for the first speaker comprises:

classifying some or all of the tagged clustered segments to identify a set of common speaker Gaussian mixture models (GMMs);

constructing a first super-GMM for the set of common speaker GMMs; and

constructing a second super-GMM for a set of generic speaker GMMs, wherein the acoustic signature comprises the first super-GMM and the second super-GMM.

18. The system according to claim 17 , further comprising using the metadata to select the linguistic model to create the speech to text transcription.

19. The system according to claim 17 , further comprising filtering out non-speech frames by evaluating envelope energy level of respective frames and comparing the envelope energy to a threshold energy above which a frame includes an utterance.

20. The system according to claim 17 , further comprising assigning start time frames to the word sequences, wherein the clustering is further based on respective time lengths of the word sequences.