This paper analyses the performance of different types of Deep Neural Networks to jointly estimate age and identify gender from speech, to be applied in Interactive Voice Response systems available in call centres. Deep Neural Networks are used, because they have recently demonstrated discriminative and representation capabilities in a wide range of applications, including speech processing problems based on feature extraction and selection. Networks with different sizes are analysed to obtain information on how performance depends on the network architecture and the number of free parameters. The speech corpus used for the experiments is Mozilla’s Common Voice dataset, an open and crowdsourced speech corpus. The results are really good for gender classification, independently of the type of neural network, but improve with the network size. Regarding the classification by age groups, the combination of convolutional neural networks and temporal neural networks seems to be the best option among the analysed, and again, the larger the size of the network, the better the results. The results are promising for use in IVR systems, with the best systems achieving a gender identification error of less than 2% and a classification error by age group of less than 20%.

The overall aim of our research is to develop methods for a monitoring system to be used at neonatal intensive care units. When monitoring a baby, a range of different types of background activity needs to be considered. In this work, we have developed a scheme for automatic classification of background EEG activity in newborn babies. EEG from six full-term babies who were displaying a burst suppression pattern while suffering from the after-effects of asphyxia during birth was included along with EEG from 20 full-term healthy newborn babies. The signals from the healthy babies were divided into four behavioural states: active awake, quiet awake, active sleep and quiet sleep. By using a number of features extracted from the EEG together with Fisher's linear discriminant classifier we have managed to achieve 100% correct classification when separating burst suppression EEG from all four healthy EEG types and 93% true positive classification when separating quiet sleep from the other types. The other three sleep stages could not be classified. When the pathological burst suppression pattern was detected, the analysis was taken one step further and the signal was segmented into burst and suppression, allowing clinically relevant parameters such as suppression length and burst suppression ratio to be calculated. The segmentation of the burst suppression EEG works well, with a probability of error around 4%.

Determining the stress level of a subject in real time could be of special interest in certain professional activities to allow the monitoring of soldiers, pilots, emergency personnel and other professionals responsible for human lives. Assessment of current mental fitness for executing a task at hand might avoid unnecessary risks. To obtain this knowledge, two physiological measurements were recorded in this work using customized non-invasive wearable instrumentation that measures electrocardiogram (ECG) and thoracic electrical bioimpedance (TEB) signals. The relevant information from each measurement is extracted via evaluation of a reduced set of selected features. These features are primarily obtained from filtered and processed versions of the raw time measurements with calculations of certain statistical and descriptive parameters. Selection of the reduced set of features was performed using genetic algorithms, thus constraining the computational cost of the real-time implementation. Different classification approaches have been studied, but neural networks were chosen for this investigation because they represent a good tradeoff between the intelligence of the solution and computational complexity. Three different application scenarios were considered. In the first scenario, the proposed system is capable of distinguishing among different types of activity with a 21.2% probability error, for activities coded as neutral, emotional, mental and physical. In the second scenario, the proposed solution distinguishes among the three different emotional states of neutral, sadness and disgust, with a probability error of 4.8%. In the third scenario, the system is able to distinguish between low mental load and mental overload with a probability error of 32.3%. The computational cost was calculated, and the solution was implemented in commercially available Android-based smartphones. The results indicate that execution of such a monitoring solution is negligible compared to the nominal computational load of current smartphones.

Studies of sensory capacities in past life forms have offered new insights into their adaptations and lifeways. Audition is particularly amenable to study in fossils because it is strongly related to physical properties that can be approached through their skeletal structures. We have studied the anatomy of the outer and middle ear in the early hominin taxa Australopithecus africanus and Paranthropus robustus and estimated their auditory capacities. Compared with chimpanzees, the early hominin taxa are derived toward modern humans in their slightly shorter and wider external auditory canal, smaller tympanic membrane, and lower malleus/incus lever ratio, but they remain primitive in the small size of their stapes footplate. Compared with chimpanzees, both early hominin taxa show a heightened sensitivity to frequencies between 1.5 and 3.5 kHz and an occupied band of maximum sensitivity that is shifted toward slightly higher frequencies. The results have implications for sensory ecology and communication, and suggest that the early hominin auditory pattern may have facilitated an increased emphasis on short-range vocal communication in open habitats.

This paper proposes the application of Fisher linear discriminants to the problem of speech/music classification. Fisher linear discriminants can classify between two different classes, and are based on the calculation of some kind of centroid for the training data corresponding with each one of these classes. Based on that information a linear boundary is established, which will be used for the classification process. Some results will be given demonstrating the superior behavior of this classification algorithm compared with the well-known K-nearest neighbor algorithm. It will also be demonstrated that it is possible to obtain very good results in terms of probability of error using only one feature extracted from the audio signal, being thus possible to reduce the complexity of this kind of systems in order to implement them in real-time