No abstract

Abstract. In this paper we present a novel method for adaptation of a multi-layer perceptron neural network (MLP ANN). Nowadays, the adaptation of the ANN is usually done as an incremental retraining either of a subset or the complete set of the ANN parameters. However, since sometimes the amount of the adaptation data is quite small, there is a fundamental drawback of such approach -during retraining, the network parameters can be easily overfitted to the new data. There certainly are techniques that can help overcome this problem (earlystopping, cross-validation), however application of such techniques leads to more complex and possibly more data hungry training procedure. The proposed method approaches the problem from a different perspective. We use the fact that in many cases we have an additional knowledge about the problem. Such additional knowledge can be used to limit the dimensionality of the adaptation problem. We applied the proposed method on speaker adaptation of a phoneme recognizer based on traps (Temporal Patterns) parameters. We exploited the fact that the employed traps parameters are constructed using log-outputs of mel-filter bank and by virtue of reformulating the first layer weight matrix adaptation problem as a mel-filter bank output adaptation problem, we were able to significantly limit the number of free variables. Adaptation using the proposed method resulted in a substantial improvement of phoneme recognizer accuracy.

In this paper, we present the system developed by the team from the New Technologies for the Information Society (NTIS) research center of the University of West Bohemia, for the First DIHARD Speech Diarization Challenge. The base of our system follows the currently-standard approach of segmentation, i-vector extraction, clustering, and resegmentation. Here, we describe the modifications to the system which allowed us to apply it to data from a range of different domains. The main contribution to our achievement is a Neural Network (NN) based domain classifier, which categorizes each conversation into one of the ten domains present in the development set. This classification determines the specific system configuration, such as the expected number of speakers and the stopping criterion for the hierarchical clustering. At the time of writing of this abstract, our best submission achieves a DER of 26.90% and an MI of 8.34 bits on the evaluation set (gold speech/nonspeech segmentation).

Abstract. The main goal of this paper is to explore the methods of genderdependent acoustic modeling that would take the possibly of imperfect function of a gender detector into consideration. Such methods will be beneficial in realtime recognition tasks (eg. real-time subtitling of meetings) when the automatic gender detection is delayed or incorrect. The goal is to minimize an impact to the correct function of the recognizer. The paper also describes a technique of unsupervised splitting of training data, which can improve gender-dependent acoustic models trained on the basis of manual markers (male/female). The idea of this approach is grounded on the fact that a significant amount of "masculine" female and "feminine" male voices occurring in training corpora and also on frequent errors in manual markers.

No abstract