Back to news overview

OpenVokaturi 1.1: less sadness

29 April 2016

The classification is now 61.4 percent. This is lower than in version 1.0, but it signals an improvement. The classification model now has equal cue standard deviations within every emotion class.

In version 1.0, the standard deviation of pitSlo was approximately 9.1 semitones per second for neutrality and 18.4 for sadness, leading to a high probability for sadness for recordings that were on the far super-neutral side of the neutralitysadness continuum. Having such discontiguous categories is a well-known problem in linear discriminant analysis and Gaussian mixture techniques. In 1.1 we therefore assume that the covariance matrices are identical for all emotions, and we compute the resulting single covariance matrix by pooling all within-emotion covariance matrices; the standard deviation of pitSlo therefore becomes approximately 12.9 semitones per second, independently from the emotion class.

We can say that the model now uses many fewer adjustable parameters, and that it is more robust, i.e. the mistakes it makes are less surprising.