Abstract 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%.

Gunshot acoustic analysis is a field with many practical applications, but due to the multitude of factors involved in the generation of the acoustic signature of firearms, it is not a trivial task. The main problem arises with the strong spatial dependence shown by the recorded waveforms even when dealing with the same weapon. However, this can be lessen by using a spatially diverse receiver such as a wireless acoustic sensor network. In this work, we address multichannel acoustic weapon classification using spatial information and a novel decision fusion rule based on it. We propose a fusion rule based on maximum likelihood estimation that takes advantage of diverse classifier ensembles to improve upon classic decision fusion techniques. Classifier diversity comes from a spatial segmentation that is performed locally at each node. The same segmentation is also used to improve the accuracy of the local classification by means of a divide and conquer approach.

We propose a new acoustic self-localization and orientation estimation algorithm for smartphones networks composed of commercial off-the-shelf devices equipped with two microphones and a speaker. Each smartphone acts as an acoustic transceiver, which emits and receives acoustic signals. Node locations are found by combining estimates of the range and direction of arrival (DoA) between node pairs using a maximum likelihood (ML) estimator. A tailored optimization algorithm is proposed to simultaneously solve the DoA uncertainty problem that arises from the use of only 2 microphones per node and obtain the azimuthal orientation of each node without requiring an electronic compass.

A binaural speech enhancement algorithm that combines superdirective beamforming with time-frequency (TF) masking is proposed. Supervised machine learning is used to design a speech/noise classifier that estimates the ideal binary mask (IBM), which is further softened to reduce musical noise. The method is energy-efficient in two ways: the computational complexity is limited and the wireless data transmission optimized. The experimental work demonstrates the ability of the method to increase the intelligibility of speech corrupted by different types of noise in low SNR scenarios.

Indoor localization of multiple speech sources in wireless acoustic sensor networks (WASNs) is an open and interesting problem with many practical applications, but the presence of noise and reverberations complicates the problem. In this paper, a distributed algorithm for multiple DOA estimation of speech sources in WASNs is presented. The method exploits the sparsity of speech sources in the time-frequency domain to obtain DOA estimations locally in each node of the network. The DOA estimations of different nodes are further combined to increase the accuracy of the local DOA estimations. Since the local DOAs are estimated using only the microphones of the same node, the synchronization between input channels and localization of the microphones from different nodes are not an issue.

The solution of speech related problems such as source location or separation relies on a prior estimation of the number of sources. In this paper we propose a method for speech source enumeration based on the different relative delays that sources at different locations register at two microphones. The Probability Density Function (PDF) of the estimated delays exhibits peaks associated with each source. The Minimum Description Length (MDL) criterion is applied to the prediction error of a linear model fitted to the delay estimates. The method is validated for the estimation of different number of sources and different mixtures.

Intelligent vehicles provide several speech-based advanced features, which success rely on the robustness of the speech acquired by the vehicle microphone system. This speech can be enhanced with a microphone array by means of spatial filtering, which design needs a reliable noise model to guarantee good filtering performance. Traditionally, it has been assumed that this noise is diffuse, but different measurements in real scenarios revel that the coherence of the noise is high to consider that is completely diffuse. In this paper, we suggest that the major contribution of the noise in a car can be reduced to a finite number of uncorrelated noise sources. We propose a searching algorithm to identify the position of the most uncorrelated sources, obtaining also their relative energy. This model allows to generate reliable synthetic noise based on real measurements, which can be very useful in the design of spatial filters.

In this paper, the alignment of noisy high-resolution radar signals using the maximum position method is studied. The relationship between the shift estimation and the signal-to-noise ratio is considered. As a result, two analytical expressions are obtained that approximate the root-mean-square error of the difference in the shift estimation with and without noise. These two expressions allow us to improve the understanding of the sensitivity to noise of the Maximum Position alignment method.

Hearing loss may disqualify many people from leading a normal life, though the majority do not make use of hearing aids. This is because most hearing aids on the market cannot automatically adapt to the changing acoustical environment the user faces daily. This paper focuses on the development of an automatic sound classifier for digital hearing aids that aims to enhance listening comprehension when the user goes from one sound environment to another. Given the strong complexity constraints of these devices, reducing the number of signal-describing features which feed the automatic classifier is of great importance and becomes a challenging topic. Thus, the use of genetic algorithms with restricted search is explored for the mentioned feature selection. In an effort to evaluate its performance, the algorithm is compared with a standard unconstrained genetic algorithm and with sequential methods. The restricted search driven by the implemented genetic algorithm performs better than both the sequential methods and unconstrained genetic algorithms. It thus allows a subset of signal-describing features with lower cardinality to be selected. This may permit these selected features to be programmed on the digital signal processor that the hearing aid is based on, and to make efficient use of its limited computational facilities.

In this letter, we propose a novel matching pursuit-based method for transient modeling with application to parametric audio coding. The overcomplete dictionary for the matching pursuit is composed of wavelet functions that implement a wavelet-packet filter bank. The proposed transient modeling method is suitable to be integrated into a parametric audio coder based on the three-part model of sines, transients, and noise (STN model). Comparative analysis between wavelet and exponentially damped sinusoidal functions are shown in experimental results. The mean-squared-error performance of the proposed approach is better than that obtained with damped sinusoids.

Outlines an algorithm for searching minimum bit rate wavelet packet representations in a high quality wavelet-based perceptual audio coder. The algorithm applies wavelet tree and prototype filter adaptation to a perceptual cost function. To achieve such adaptive structure, a periodized wavelet packet transform is performed for each audio frame. After the transform, the encoder employs scalar adaptive quantization, controlled by a psycho-acoustic model in the wavelet domain, followed by entropy coding and bit packing. An adaptive Huffman algorithm that assumes Laplacian distributions for wavelet coefficients is used for entropy coding. A relevant feature of our coder is the method to translate the psycho-acoustic information to the wavelet domain, which allows using orthonormal wavelets of any compact support. Experimental results indicate that our coder ensures 'transparent' coding of monophonic CD audio signals at bit rates lower than 64 kb/s for all the test signals.