The most widely spread measure of performance, accuracy, suffers from a paradox: predictive models with a given level of accuracy may have greater predictive power than models with higher accuracy. Despite optimizing classification error rate, high accuracy models may fail to capture crucial information transfer in the classification task. We present evidence of this behavior by means of a combinatorial analysis where every possible contingency matrix of 2, 3 and 4 classes classifiers are depicted on the entropy triangle, a more reliable information-theoretic tool for classification assessment. Motivated by this, we develop from first principles a measure of classification performance that takes into consideration the information learned by classifiers. We are then able to obtain the entropy-modulated accuracy (EMA), a pessimistic estimate of the expected accuracy with the influence of the input distribution factored out, and the normalized information transfer factor (NIT), a measure of how efficient is the transmission of information from the input to the output set of classes. The EMA is a more natural measure of classification performance than accuracy when the heuristic to maximize is the transfer of information through the classifier instead of classification error count. The NIT factor measures the effectiveness of the learning process in classifiers and also makes it harder for them to “cheat” using techniques like specialization, while also promoting the interpretability of results. Their use is demonstrated in a mind reading task competition that aims at decoding the identity of a video stimulus based on magnetoencephalography recordings. We show how the EMA and the NIT factor reject rankings based in accuracy, choosing more meaningful and interpretable classifiers.

The main research motivation of this paper is the fight against gender-based violence and achieving gender equality from a technological perspective. The solution proposed in this work goes beyond currently existing panic buttons, needing to be manually operated by the victims under difficult circumstances. Instead, Bindi, our end-to-end autonomous multimodal system, relies on artificial intelligence methods to automatically identify violent situations, based on detecting fear-related emotions, and trigger a protection protocol, if necessary. To this end, Bindi integrates modern state-of-the-art technologies, such as the Internet of Bodies, affective computing, and cyber-physical systems, leveraging i) affective IoT with auditory and physiological commercial off-the-shelf smart sensors embedded in wearable devices, ii) hierarchical multisensorial information fusion, and iii) the edge-fog-cloud IoT architecture. This solution is evaluated using our own dataset named WEMAC, a very recently collected and freely-available collection of data comprising the auditory and physiological responses of 47 women to several emotions elicited by using a virtual reality environment. On this basis, this work provides an analysis of multimodal late fusion strategies to combine the physiological and speech data processing pipelines to identify the best intelligence engine strategy for Bindi. In particular, the best data fusion strategy reports an overall fear classification accuracy of 63.61% for a subject-independent approach. Both a power consumption study and an audio data processing pipeline to detect violent acoustic events complement this analysis. This research is intended as an initial multimodal baseline that facilitates further work with real-life elicited fear in women.

The improved theoretical properties of Support Vector Machines with respect to other machine learning alternatives due to their max-margin training paradigm have led us to suggest them as a good technique for robust speech recognition. However, important shortcomings have had to be circumvented, the most important being the normalisation of the time duration of different realisations of the acoustic speech units. In this paper, we have compared two approaches in noisy environments: first, a hybrid HMM–SVM solution where a fixed number of frames is selected by means of an HMM segmentation and second, a normalisation kernel called Dynamic Time Alignment Kernel (DTAK) first introduced in Shimodaira et al. [Shimodaira, H., Noma, K., Nakai, M., Sagayama, S., 2001. Support vector machine with dynamic time-alignment kernel for speech recognition. In: Proc. Eurospeech, Aalborg, Denmark, pp. 1841–1844] and based on DTW (Dynamic Time Warping). Special attention has been paid to the adaptation of both alternatives to noisy environments, comparing two types of parameterisations and performing suitable feature normalisation operations. The results show that the DTA Kernel provides important advantages over the baseline HMM system in medium to bad noise conditions, also outperforming the results of the hybrid system.Publicad

In the last years, support vector machines (SVMs) have shown excellent performance in many applications, especially in the presence of noise. In particular, SVMs offer several advantages over artificial neural networks (ANNs) that have attracted the attention of the speech processing community. Nevertheless, their high computational requirements prevent them from being used in practice in automatic speech recognition (ASR), where ANNs have proven to be successful. The high complexity of SVMs in this context arises from the use of huge speech training databases with millions of samples and highly overlapped classes. This paper suggests the use of a weighted least squares (WLS) training procedure that facilitates the possibility of imposing a compact semiparametric model on the SVM, which results in a dramatic complexity reduction. Such a complexity reduction with respect to conventional SVMs, which is between two and three orders of magnitude, allows the proposed hybrid WLS-SVC/HMM system to perform real-time speech decoding on a connected-digit recognition task (SpeechDat Spanish database). The experimental evaluation of the proposed system shows encouraging performance levels in clean and noisy conditions, although further improvements are required to reach the maturity level of current context-dependent HMM-based recognizers.