Speech recognition with a new hybrid architecture combining neural networks and continuous HMM