In order to improve the noise robustness of automatic speaker recognition, many techniques on speech/feature enhancement have been explored by using deep neural networks (DNN). In this work, a DNN multi-level enhancement (DNN-ME), which consists of the stages of signal enhancement, cepstrum enhancement and i-vector enhancement, is proposed for text-independent speaker recognition. Given the fact that these enhancement methods are applied in different stages of the speaker recognition pipeline, it is worth exploring the complementary role of these methods, which benefits the understanding of the pros and cons of the enhancements of different stages. In order to use the capabilities of DNN-ME as much as possible, two kinds of methods called Cascaded DNN-ME and joint input of DNNs are studied. Weighted Gaussian mixture models (WGMMs) proposed in our previous work is also applied to further improve the model's performance. Experiments conducted on the Speakers in the Wild (SITW) database have shown that DNN-ME demonstrated significant superiority over the systems with only a single enhancement for noise robust speaker recognition. Compared with the i-vector baseline, the equal error rate (EER) was reduced from 5.75 to 4.01.

Unsupervised spoken unit discovery or zero-source speech recognition is an emerging research topic which is important for spoken document analysis of languages or dialects with little human annotation. In this paper, we extend our earlier joint training framework for unsupervised learning of discrete density HMM to continuous density HMM (CDHMM) and apply it to spoken unit discovery. In the proposed recipe, we first cluster a group of Gaussians which then act as initializations to the joint training framework of nonnegative matrix factorization and semi-continuous density HMM (SCDHMM). In SCDHMM, all the hidden states share the same group of Gaussians but with different mixture weights. A CDHMM is subsequently constructed by tying the top-N activated Gaussians to each hidden state. Baum-Welch training is finally conducted to update the parameters of the Gaussians, mixture weights and HMM transition probabilities. Experiments were conducted on word discovery from TIDIGITS and phone discovery from TIMIT. For TIDIGITS, units were modeled by 10 states which turn out to be strongly related to words; while for TIMIT, units were modeled by 3 states which are likely to be phonemes.