The features used for training contain the bitmap-vector and the additional features as described above (altogether a 12 dimensional feature vector), the codebook sizes are optimized for each experiment, whereas the sizes of the neural VQ for hybrid HMMs and the corresponding k-means VQ are equal.