The automatic classification of content is an essential requirement for multimedia applications. Present research for audio-based classifiers uses short- and long-term analysis of signals, with temporal and spectral features. In our prior study, we presented an approach to classify streaming and local content, in real-time and with low latency, using synthetically-derived metadata features based on fixed class-conditional distributions. The three-class conditional distribution parameters were set a priori based on public information. In this paper, we extend the approach to optimizing the class condition distribution parameters, the neural network hyperparameters, and the size of the synthetic metadata training set using a combination of Bayesian optimization and by invoking the VC-dimensionality. We demonstrate that the resulting classifier is then able to improve classification accuracy by at least 10% with a modest increase in the computational complexity (in terms of multiply-adds) and decision time.

Automatic content classification is an essential problem for multimedia applications. In this paper, we present a metadata synthesis approach, using real-world distributions, and a machine learning (ML) model for classifying content for local and streaming sources. The synthetic metadata is used to train a neural network. The neural network is then able to classify content (e.g., movies, voice, and music) based on real metadata from YouTube and local content. We demonstrate high accuracy ~ 90% with low latency (viz., on an average 7 ms). We also identified the contribution of each synthetic metadata feature towards the classification performance on real-world content.

Data transforms, parameter re-normalization, and activation functions have gained significant attention in the neural network community over the past several years for improving convergence speed. The results in the literature are for computer vision applications, with batch-normalization (BN) and the Rectified Linear Unit (ReLU) activation attracting attention. In this paper, we present a new approach in data transformation in the context of regression during the synthesis of Head-related Transfer Functions (HRTFs) in the field of audio. The encoding technique whitens the real-valued input data delivered to the first hidden layer of a fully-connected neural network (FCNN) thereby providing the training speedup. The experimental results demonstrate, in a statistically significant way, that the presented data encoding approach outperforms other forms of normalization in terms of convergence speed, lower mean-square error, and robustness to network parameter initialization. Towards this, we used some popular first-and second-order gradient techniques such as scaled conjugate gradient, Extreme Learning Machine (ELM), and stochastic gradient descent with momentum and batch normalization. The improvements, as shown through t-SNE based depiction and analysis on the input covariance matrix, confirm the reduction in the condition number of the input covariance matrix (a process similar to whitening).

In this paper, a new method, using the Sammon map, is proposed for visualizing room responses and room response equalization. Specifically, the map provides an important perspective on the formation of clusters of acoustic room responses when cluster validity measures are unable to clearly identify the number of clusters. Additionally, the Sammon map can also be used for displaying the uniformity of the equalized responses at multiple positions. Furthermore, the map does not impose significant computational requirements, when the number of measured responses is reasonably small, and hence can be readily implemented for analyzing multiple position (viz., multiple listener) room equalization performance

Low frequency room acoustic equalization is an extremely challenging problem to solve with realizable digital equalization filters. In this paper we propose a cascaded FIR solution, using a pattern recognition technique, for multiposition room equalization at low frequencies. The cascaded FIR solution allows for a real-time implementation of realizable filters whereas the pattern recognition approach captures the variabilities in room acoustics between different listening positions for multiposition equalization. We compare the results with the popular warping with low order spectral modeling approach, as well as with the multirate approach. The proposed method outperforms both the warping based and the multirate methods as can be evidenced through published data

Given a multichannel loudspeaker system, the selection of the crossover frequency between the sub-woofer and the satellite speakers is important for accurate (i.e., distortion-free), reproduction of playback sound. Presently, many home-theater systems have selectable crossover frequencies which are either set by the consumer through listening tests. Alternatively, if the loudspeakers are industry certified, the crossover frequency is set at 80 Hz. A desirable feature is that, besides distortion-free sound output from the individual subwoofer and the satellite speakers, the combined sub-woofer and satellite room acoustical response should exhibit negligible variations around the selected crossover frequency. In this paper, we present an automatic crossover frequency selection algorithm based on an objective measure (viz., the spectral deviation measure) for multichannel home-theater applications that allows better control of the combined subwoofer-satellite response, thereby significantly improving audio quality.

Traditionally, room response equalization is performed to improve sound quality at a given listener in applications ranging from automobile, home-theater, movie theater, multimedia education in classrooms. However, room responses vary with source and listener positions. Hence, in a multiple listener environment, equalization may be performed through spatial averaging of magnitude responses at locations of interest (e.g., in movie theater equalization). However, the performance of averaging based equalization, at the listeners, may be affected when listener positions change, or due to mismatch between microphone and listener positions (i.e., displacement effects). In this paper, we present a statistical approach to map displacement effects to a magnitude response averaging equalization performance metric. The results indicate that, for the analyzed listener configurations, the zone of equalization depends on, (i) distance of microphones/listeners from a source, (ii) the listener arrangement, and (iii) the source signal spectral composition.

Selectively canceling signals at specific locations within an acoustical environment with multiple listeners is of significant importance for home theater, automobile, teleconferencing, office, industrial, and other applications. The traditional noise cancellation approach is impractical for such applications because it requires secondary sources to "anti-phase" the primary source, or sensors to be placed on the listeners. In this paper we propose an alternative method for signal cancellation by pre-processing the acoustical signal with a filter known as the eigenfilter , . We examine the theoretical properties of such filters, and investigate the performance (gain) and tradeoff issues such as, spectral distortion. Sensitivity of the performance as a function of the room impulse response duration (reverberation) modeled in the eigenfilter is also investigated.

We propose a novel method of visualizing room equalization at multiple locations using Sammon mapping. The benefits of this method are: (i) simplicity in implementation; (ii) negligible computational requirements to implement the map; (iii) realistic evaluation of equalization performance for our cluster based method in room response equalization at multiple listener locations. Furthermore, the mapping provides an interesting perspective on (i) how the clusters are formed, (ii) how the centroids are determined, and (iii) the result of combining the centroids in a specific manner (to form the final prototype which is then inverted to form the equalization filter). We also compare the results obtained from using the Sammon map to the results obtained by using the spectral deviation measure.

A new neuron activation function based on a property found in physical systems--hysteresis--is proposed. We incorporate this neuron activation in a fully connected dynamical system to form the hysteretic Hopfield neural network (HHNN). We then present an analog implementation of this architecture and its associated dynamical equation and energy function.We proceed to prove Lyapunov stability for this new model, and then solve a combinatorial optimization problem (i.e., the N-queen problem) using this network. We demonstrate the advantages of hysteresis by showing increased frequency of convergence to a solution, when the parameters associated with the activation function are varied.

Microcode optimization is an NP-complete combinatorial optimization problem. This paper proposes a new method based on the Hopfield neural network for optimizing the wordwidth in the control memory of a microprogrammed digital computer. We present two methodologies, viz., the maximum clique approach, and a cost function based method to minimize an objective function. The maximum clique approach albeit being near O(1) in complexity, is limited in its use for small problem sizes, since it only partitions the data based on the compatibility between the microoperations, and does not minimize the cost function. We thereby use this approach to condition the data initially (to form compatibility classes), and then use the proposed second method to optimize on the cost function. The latter method is then able to discover better solutions than other schemes for the benchmark data set.