Investigation of multilingual deep neural networks for spoken term detection

The development of high-performance speech processing systems for low-resource languages is a challenging area. One approach to address the lack of resources is to make use of data from multiple languages. A popular direction in recent years is to use bottleneck features, or hybrid systems, trained on multilingual data for speech-to-text (STT) systems. This paper presents an investigation into the application of these multilingual approaches to spoken term detection. Experiments were run using the IARPA Babel limited language pack corpora (∼10 hours/language) with 4 languages for initial multilingual system development and an additional held-out target language. STT gains achieved through using multilingual bottleneck features in a Tandem configuration are shown to also apply to keyword search (KWS). Further improvements in both STT and KWS were observed by incorporating language questions into the Tandem GMM-HMM decision trees for the training set languages. Adapted hybrid systems performed slightly worse on average than the adapted Tandem systems. A language independent acoustic model test on the target language showed that retraining or adapting of the acoustic models to the target language is currently minimally needed to achieve reasonable performance. © 2013 IEEE.

Free-Space Interconnect for Optical Switch and Analogue Neural Network

In this paper, the architecture of a vector-matrix multiplier (MVM) is simulated. The optical design can be made compact by the use of GRIN lenses for the optical fan-in. The intended application area was in storage area networks (SANs) but the concept can be applied to a neural network. © 2011 Allerton Press, Inc.

Pixelated liquid-crystal light valve for neural network application.

The liquid-crystal light valve (LCLV) is a useful component for performing integration, thresholding, and gain functions in optical neural networks. Integration of the neural activation channels is implemented by pixelation of the LCLV, with use of a structured metallic layer between the photoconductor and the liquid-crystal layer. Measurements are presented for this type of valve, examples of which were prepared for two specific neural network implementations. The valve fabrication and measurement were carried out at the State Optical Institute, St. Petersburg, Russia, and the modeling and system applications were investigated at the Institute of Microtechnology, Neuchâtel, Switzerland.

Gaussian Process Regression Networks

We introduce a new regression framework, Gaussian process regression networks (GPRN), which combines the structural properties of Bayesian neural networks with the non-parametric flexibility of Gaussian processes. This model accommodates input dependent signal and noise correlations between multiple response variables, input dependent length-scales and amplitudes, and heavy-tailed predictive distributions. We derive both efficient Markov chain Monte Carlo and variational Bayes inference procedures for this model. We apply GPRN as a multiple output regression and multivariate volatility model, demonstrating substantially improved performance over eight popular multiple output (multi-task) Gaussian process models and three multivariate volatility models on benchmark datasets, including a 1000 dimensional gene expression dataset.

Multi-walled carbon nanotube conductivity enhancement and band gap widening via rapid pulsed thermal annealing

Herein we report on the transport characteristics of rapid pulsed vacuum-arc thermally annealed, individual and network multi-walled carbon nanotubes. Substantially reduced defect densities (by at least an order of magnitude), measured by micro-Raman spectroscopy, and were achieved by partial reconstruction of the bamboo-type defects during thermal pulsing compared with more traditional single-pulse thermal annealing. Rapid pulsed annealed processed networks and individual multi-walled nanotubes showed a consistent increase in conductivity (of over a factor of five at room temperature), attributed to the reduced number density of resistive axial interfaces and, in the case of network samples, the possible formation of structural bonds between crossed nanotubes. Compared to the highly defective as-grown nanotubes, the pulsed annealed samples exhibited reduced temperature sensitivity in their transport characteristics signifying the dominance of scattering events from structural defects. Transport measurements in the annealed multi-walled nanotubes deviated from linear Ohmic, typically metallic, behavior to an increasingly semiconducting-like behavior attributed to thermally induced axial strains. Rapid pulsed annealed networks had an estimated band gap of 11.26 meV (as-grown; 6.17 meV), and this observed band gap enhancement was inherently more pronounced for individual nanotubes compared with the networks most likely attributed to mechanical pinning effect of the probing electrodes which possibly amplifies the strain induced band gap. In all instances the estimated room temperature band gaps increased by a factor of two. The gating performance of back-gated thin-film transistor structures verified that the observed weak semiconductivity (p-type) inferred from the transport characteristic at room temperature. © 2014 Copyright Taylor & Francis Group, LLC.