Short-duration speaker modelling with phone adaptive training