Adapting Hybrid ANN/HMM to Speech Variations