Structure and Base Analysis of Receptive Field Neural Networks in a Character Recognition Task

This paper explores extensions and restrictions of shallow convolutional neural networks with fixed kernels trained with a limited number of training samples. We extend the work recently done in research on Receptive Field Neural Networks (RFNN) and show their behaviour using different bases and step-by-step changes within the network architecture. To ensure the reproducibility of the results, we simplified the baseline RFNN architecture to a single-layer CNN network and introduced a deterministic methodology for RFNN training and evaluation. This methodology enabled us to evaluate the significance of changes using the (recently widely used in neural networks) Bayesian comparison. The results indicate that a change in the base may have less of an effect on the results than re-training using another seed. We show that the simplified network with tested bases has similar performance to the chosen baseline RFNN architecture. The data also show the positive impact of energy normalization of used filters, which improves the classification accuracy, even when using randomly initialized filters.

Frequency, Time, Representation and Modeling Aspects for Major Speech and Audio Processing Applications

There are many speech and audio processing applications and their number is growing. They may cover a wide range of tasks, each having different requirements on the processed speech or audio signals and, therefore, indirectly, on the audio sensors as well. This article reports on tests and evaluation of the effect of basic physical properties of speech and audio signals on the recognition accuracy of major speech/audio processing applications, i.e., speech recognition, speaker recognition, speech emotion recognition, and audio event recognition. A particular focus is on frequency ranges, time intervals, a precision of representation (quantization), and complexities of models suitable for each class of applications. Using domain-specific datasets, eligible feature extraction methods and complex neural network models, it was possible to test and evaluate the effect of basic speech and audio signal properties on the achieved accuracies for each group of applications. The tests confirmed that the basic parameters do affect the overall performance and, moreover, this effect is domain-dependent. Therefore, accurate knowledge of the extent of these effects can be valuable for system designers when selecting appropriate hardware, sensors, architecture, and software for a particular application, especially in the case of limited resources.

On the Speech Properties and Feature Extraction Methods in Speech Emotion Recognition

Many speech emotion recognition systems have been designed using different features and classification methods. Still, there is a lack of knowledge and reasoning regarding the underlying speech characteristics and processing, i.e., how basic characteristics, methods, and settings affect the accuracy, to what extent, etc. This study is to extend physical perspective on speech emotion recognition by analyzing basic speech characteristics and modeling methods, e.g., time characteristics (segmentation, window types, and classification regions—lengths and overlaps), frequency ranges, frequency scales, processing of whole speech (spectrograms), vocal tract (filter banks, linear prediction coefficient (LPC) modeling), and excitation (inverse LPC filtering) signals, magnitude and phase manipulations, cepstral features, etc. In the evaluation phase the state-of-the-art classification method and rigorous statistical tests were applied, namely N-fold cross validation, paired t-test, rank, and Pearson correlations. The results revealed several settings in a 75% accuracy range (seven emotions). The most successful methods were based on vocal tract features using psychoacoustic filter banks covering the 0–8 kHz frequency range. Well scoring are also spectrograms carrying vocal tract and excitation information. It was found that even basic processing like pre-emphasis, segmentation, magnitude modifications, etc., can dramatically affect the results. Most findings are robust by exhibiting strong correlations across tested databases.