Improvements to the speech activity detection model.

Improvements to the speech activity detection model.

The analyser outputs 2 new estimates
* frame-level estimate corresponding to the raw speech/non speech estimate, filtered with dilatation and erosion procedures, remove some outliers
* speech-non speech segments, inferred from the dilatation erosion/procedure

The analyser can be initialised with 3 new options:
* size of the window used for the dilatation and erosion procedures
* threshold for the decision of speech VS non speech
* bounds for the raw estimates (usefull for plotting)