arXiv subjects

Venkateshan Kannan

Publications and source records attributed to Venkateshan Kannan.

7 recordsLinked to original sources

arXiv2021

Attention based end to end Speech Recognition for Voice Search in Hindi and English

We describe here our work with automatic speech recognition (ASR) in the context of voice search functionality on the Flipkart e-Commerce platform. Starting with the deep learning architecture of Listen-Attend-Spell (LAS), we build upon and expand the model design and attention mechanisms to incorporate innovative approaches including multi-objective training, multi-pass training, and external rescoring using language models and phoneme based losses. We report a relative WER improvement of 15.7% on top of state-of-the-art LAS models using these modifications. Overall, we report an improvement of 36.9% over the phoneme-CTC system. The paper also provides an overview of different components that can be tuned in a LAS-based system.

eess.AS↗

arXiv2017

Characterization and reduction of variability in selection based on effect-size using association measures in cohort study of heterogeneous diseases

Cohort studies employ pairwise measures of association to quantify dependencies among conditions and exposures. To reliably use these measures to draw conclusions about the underlying association strengths requires that the measures be robust and unbiased. These considerations assume greater significance when applied to disease networks, where associations among heterogeneous pairs of diseases are ranked. Using disease diagnoses data from a large cohort of 5.5 million individuals, we develop a comprehensive methodology to characterize the bias of standard association measures like relative risk and $ϕ$ correlation. To overcome these biases, we devise a novel measure based on a stochastic model for disease development. The new measure is demonstrated to have the least overall bias and hence would be most suitable for application to heterogeneous disease cohorts.

q-bio.QM↗

arXiv2014

On the Theory and Algorithm for rigorous discretization in applications of Information Theory

We identify fundamental issues with discretization when estimating information-theoretic quantities in the analysis of data. These difficulties are theoretical in nature and arise with discrete datasets carrying significant implications for the corresponding claims and results. Here we describe the origins of the methodological problems, and provide a clear illustration of their impact with the example of biological network reconstruction. We propose an algorithm (shared information metric) that corrects for the biases and the resulting improved performance of the algorithm demonstrates the need to take due consideration of this issue in different contexts.

q-bio.QM↗

arXiv2013

Harmonic Systems With Bulk Noises

We consider a harmonic chain in contact with thermal reservoirs at different temperatures and subject to bulk noises of different types: velocity flips or self-consistent reservoirs. While both systems have the same covariances in the nonequilibrium stationary state (NESS) the measures are very different. We study hydrodynamical scaling, large deviations, fluctuations, and long range correlations in both systems. Some of our results extend to higher dimensions.

cond-mat.stat-mech↗

arXiv2012

Non-equilibrium stationary state of a harmonic crystal with alternating masses

We analyze the non-equilibrium steady states (NESS) of a one dimensional harmonic chain of $N$ atoms with alternating masses connected to heat reservoirs at unequal temperatures. We find that the temperature profile defined through the local kinetic energy $T(j) \equiv { }/{m_j}$, oscillates with period two in the bulk of the system. Depending on boundary conditions, either the heavier or the lighter particles in the bulk are hotter. We obtain exact expressions for the bulk temperature profile and steady state current in the limit $N \rightarrow \infty$. These depend on whether $N$ is odd or even. We also study similar temperature oscillations in the NESS of systems with noise in the dynamics. These die out as $N \rightarrow \infty$.

math-ph↗

arXiv2011

Nonequilibrium Stationary States of Harmonic Chains with Bulk Noises

We consider a chain composed of $N$ coupled harmonic oscillators in contact with heat baths at temperature $T_\ell$ and $T_r$ at sites 1 and $N$ respectively. The oscillators are also subjected to non-momentum conserving bulk stochastic noises. These make the heat conductivity satisfy Fourier's law. Here we describe some new results about the hydrodynamical equations for typical macroscopic energy and displacement profiles, as well as their fluctuations and large deviations, in two simple models of this type.

cond-mat.stat-mech↗

arXiv2011

Heat conduction in disordered harmonic lattices with energy conserving noise

We study heat conduction in a harmonic crystal whose bulk dynamics is supplemented by random reversals (flips) of the velocity of each particle at a rate $λ$. The system is maintained in a nonequilibrium stationary state(NESS) by contacts with Langevin reservoirs at different temperatures. We show that the one-body and pair correlations in this system are the same (after an appropriate mapping of parameters) as those obtained for a model with self-consistent reservoirs. This is true both for the case of equal and random(quenched) masses. While the heat conductivity in the NESS of the ordered system is known explicitly, much less is known about the random mass case. Here we investigate the random system, with velocity flips. We improve the bounds on the Green-Kubo conductivity obtained by C.Bernardin. The conductivity of the 1D system is then studied both numerically and analytically. This sheds some light on the effect of noise on the transport properties of systems with localized states caused by quenched disorder.

cond-mat.stat-mech↗