On MLP-based Posterior Features for Template-based ASR

We investigate the invariance of posterior features estimated using MLP trained on auxiliary corpus towards different data condition and different distance measures for matching posterior features in the context of template-based ASR. Through ASR studies on isolated word recognition task we show that posterior features estimated using MLP trained on auxiliary corpus with out any kind of adaptation can achieve comparable or better performance when compared to the case where the MLP is trained on the corpus same as that of the test set. We also show that local scores, weighted symmetric KL-divergence and Bhattacharya distance yield better systems compared to Hellinger distance, cosine angle, L1-norm, L2-norm, dot product, and cross entropy.