Efficient speech recognition using subvector quantization and discrete-mixture HMMS