Speaker adaptation : modelling variabilities