In this paper, we combine Hidden Markov Models (HMMs) with i-vector extractors to address the problem of text-dependent speaker recognition with random digit strings. We employ digit-specific HMMs to segment the utterances into digits, to perform frame alignment to HMM states and to extract Baum-Welch statistics. By making use of the natural partition of input features into digits, we train digit-specific i-vector extractors on top of each HMM and we extract well-localized i-vectors, each modelling merely the phonetic content corresponding to a single digit. We then examine ways to perform channel and uncertainty compensation, and we propose a novel method for using the uncertainty in the i-vector estimates. The experiments on RSR2015 part III show that the proposed method attains 1.52% and 1.77% Equal Error Rate (EER) for male and female respectively, outperforming state-of-the-art methods such as x-vectors, trained on vast amounts of data. Furthermore, these results are attained by a single system trained entirely on RSR2015, and by a simple score-normalized cosine distance. Moreover, we show that the omission of channel compensation yields only a minor degradation in performance, meaning that the system attains state-of-the-art results even without recordings from multiple handsets per speaker for training or enrolment. Similar conclusions are drawn from our experiments on the RedDots corpus, where the same method is evaluated on phrases. Finally, we report results with bottleneck features and show that further improvement is attained when fusing them with spectral features.

Abstract The progressive increase of information content has recently made it necessary to create a system for automatic classification of documents. In this article, a system is presented for the categorization of multiclass Farsi documents that requires fewer training examples and can help to compensate the shortcoming of the standard training dataset. The new idea proposed in the present article is based on extending the feature vector by adding some words extracted from a thesaurus and then filtering the new feature vector by applying secondary feature selection to discard inappropriate features. In fact, a phase of secondary feature selection is applied to choose more appropriate features among the features added from a thesaurus to enhance the effect of using a thesaurus on the efficiency of the classifier. To evaluate the proposed system, a corpus is gathered from the Farsi Wikipedia website and some articles in the Hamshahri newspaper, the Roshd periodical, and the Soroush magazine. In addition to studying the role of a thesaurus and applying secondary feature selection, the effect of a various number of categories, size of the training dataset, and average number of words in the test data also are examined. As the results indicate, classification efficiency improves by applying this approach, especially when available data is not sufficient for some text categories.