System and method for real-time speaker segmentation of audio interactions

1. A method of associating a segment of a voice interaction with a speaker, the method comprising:

generating a plurality of first models by, for each of a plurality of speakers:

generating a first model used to identify the speaker based on analyzing one or more recordings of voice interactions that include the speaker, the first model comprising voice characteristics of the speaker;

while a voice interaction different than a voice interaction used to generate a model is in progress:

obtaining, by a controller, a digital representation of a portion of the voice interaction and selecting a segment therefrom;

associating, by the controller, the segment with a specific speaker of the plurality of speakers based on a first model of the plurality of first models, the first model unique to the specific speaker;

updating the first model based on the segment to produce an updated model; and

using the updated model to associate a subsequent segment of the voice interaction with the specific speaker of the plurality of speakers; and

wherein associating the segment with a speaker is performed when the identity of the speaker is unknown and comprises searching a set of models for a model matching the speaker, the matching based on an acoustic feature.

2. The method of claim 1 , comprising:

obtaining a second model related to a second speaker;

calculating for the segment a first score and a second score based on the first and second models; and

selecting to associate the segment with one of the first and second speakers based on the calculated scores.

3. The method of claim 2 , wherein the first and second scores are calculated based on calculating one or more probabilities of a respective one or more acoustic features given the first and second models.

4. The method of claim 1 , comprising iteratively and in real-time, updating the first model based on classified segments of the interaction.

5. The method of claim 1 , wherein the segment is selected based on a voice activity detection performed on at least part of the portion.

6. The method of claim 1 , comprising:

determining at least one acoustic feature in the segment;

calculating at least one probability value for the at least one acoustic feature based on the first model; and

associating the segment with the speaker if the at least one probability value meets a predefined criteria.

7. The method of claim 6 , comprising:

associating a plurality of segments with the speaker according to a respective plurality of probability values; and

updating the first model based on a subset of the segments, the subset selected based on the probability values.

8. The method of claim 1 , comprising determining at least one parameter related to the segment and selecting the first model, in real-time, based on the at least one parameter.

9. The method of claim 1 , comprising analyzing the segment to produce an analysis result and selecting the first model, in real-time, from a plurality of models based on the analysis result.

10. The method of claim 1 , comprising analyzing the segment to produce an analysis result and generating the first model, in real-time, based on the analysis result.

11. The method of claim 1 , comprising identifying the speaker by relating the updated model to a plurality of models associated with a respective plurality of known speakers.

12. The method of claim 1 , comprising continuously improving the model while the second voice interaction is in progress such that the model is better adapted to associate segments of the interaction with the speaker.

13. An article comprising a non-transitory computer-readable storage medium, having stored thereon instructions, that when executed on a computer, cause the computer to:

generate a plurality of first models by, for each of a plurality of speakers:

generate a first model used to identify the speaker based on analyzing one or more recordings of voice interactions that include the speaker, the first model comprising voice characteristics of the speaker;

while a voice interaction different than a voice interaction used to generate a model is in progress:

obtain a digital representation of a portion of the voice interaction;

analyze the portion and select a segment therefrom;

associate the segment with a specific speaker of the plurality of speakers based on a first model of the plurality of first models, the first model unique to the specific speaker;

update the first model based on the segment to produce an updated model; and

use the updated model to associate a subsequent segment of the voice interaction with the specific speaker of the plurality of speakers; and

wherein associating the segment with a speaker is performed when the identity of the speaker is unknown and comprises searching a set of models for a model matching the speaker, the matching based on an acoustic feature.

14. The article of claim 13 , wherein the instructions when executed further result in:

obtaining a second model related to a second speaker;

calculating for the segment a first score and a second score based on the first and second models; and

selecting to associate the segment with one of the first and second speakers based on the calculated scores.

15. The article of claim 14 , wherein the first and second scores are calculated based on calculating one or more probabilities for a respective one or more acoustic features given the first and second models.

16. The article of claim 13 , wherein the instructions when executed further result in iteratively and in real-time, updating the first model based on segments of the interaction.

17. The article of claim 13 , wherein the segment is selected based on a voice activity detection performed on at least part of the portion.

18. The article of claim 13 , wherein the instructions when executed further result in:

determining at least one acoustic feature in the segment;

calculating at least one probability value for the at least one acoustic feature based on the first model; and

associating the segment with the speaker if the at least one probability value meets a predefined criteria.

19. The article of claim 18 , wherein the instructions when executed further result in

associating a plurality of segments with the speaker according to a respective plurality of probability values; and

updating the first model based on a subset of the segments, the subset selected based on the probability scores.

20. The article of claim 13 , wherein the instructions when executed further result in analyzing the segment to produce an analysis result and selecting the first model, in real-time, from a plurality of models based on the analysis result.

21. The article of claim 13 , wherein the instructions when executed further result in identifying the speaker by relating an initial model to a plurality of models associated with a respective plurality of known speakers.