Audio call classification and survey system
A method whereby an audio content is analyzed for speech semantic content and speech vocal content to determine audio content metrics. Additionally, the method may also incorporate the similarity of the audio content to other audio contents and weight the analyses to determine audio content metrics.
1. A method of processing a current call on one more machines, the method comprising:
receiving audio content for the current call;
generating a first feature vector from the audio content representing speech semantic content of the current call;
generating a second feature vector from the audio content representing speech vocal content of the current call;
forming a combined feature vector by concatenating the first feature vector and the second feature vector;
applying the combined feature vector as an input to a multi-modal direct call grading model to generate a plurality of direct call controls;
generating a similarity matrix for the current call from the first feature vector and the second feature vector;
applying the similarity matrix and a tree structure of feature matrices for prior calls to a distance function to generate an idiosyncratic call control and a similarity call control;
weighting the direct call controls, ideosyncratic call control, and similarity call control to form a weighted call control model, wherein weights for the direct call controls model are based on an inter-correlation of a plurality of models utilized in the multi-modal direct call grading model, the weight increasing with decreasing correlation; and
applying the weighted call control model as a feedback control to influence processing of future audio calls by the one or more machines.
2. The method of claim 1 , wherein generating and applying the similarity matrix is performed in response to a set of the prior calls as recorded comprising less than a pre-determined content file size.
3. A non-transitory computer-readable medium comprising instructions that when executed by a computer processor, result in:
generating a first feature vector from audio content of a received call, the first feature vector representing speech semantic content of the current call;
generating a second feature vector from the audio content representing speech vocal content of the current call;
forming a combined feature vector by concatenating the first feature vector and the second feature vector;
applying the combined feature vector as an input to a multi-modal direct call grading model to generate a plurality of direct call controls;
generating a similarity matrix for the current call from the first feature vector and the second feature vector;
applying the similarity matrix and a tree structure of feature matrices for prior calls to a distance function to generate an idiosyncratic call control and a similarity call control;
weighting the direct call controls, ideosyncratic call control, and similarity call control to form a weighted call control model, wherein weights for the direct call controls model are based on an inter-correlation of a plurality of models utilized in the multi-modal direct call grading model, the weight increasing with decreasing correlation; and
applying the weighted call control model as a feedback control to influence processing of future audio calls by the one or more machines.
4. The non-transitory computer-readable medium of claim 3 , wherein generating and applying the similarity matrix is performed in response to a set of the prior calls as recorded comprising less than a pre-determined content file size.