Improving out-domain PLDA speaker verification using unsupervised inter-dataset variability compensation approach

Experimental studies have found that when the state-of-the-art probabilistic linear discriminant analysis (PLDA) speaker verification systems are trained using out-domain data, it significantly affects speaker verification performance due to the mismatch between development data and evaluation data. To overcome this problem we propose a novel unsupervised inter dataset variability (IDV) compensation approach to compensate the dataset mismatch. IDV-compensated PLDA system achieves over 10% relative improvement in EER values over out-domain PLDA system by effectively compensating the mismatch between in-domain and out-domain data.

Dataset-invariant covariance normalization for out-domain PLDA speaker verification

In this paper we introduce a novel domain-invariant covariance normalization (DICN) technique to relocate both in-domain and out-domain i-vectors into a third dataset-invariant space, providing an improvement for out-domain PLDA speaker verification with a very small number of unlabelled in-domain adaptation i-vectors. By capturing the dataset variance from a global mean using both development out-domain i-vectors and limited unlabelled in-domain i-vectors, we could obtain domain- invariant representations of PLDA training data. The DICN- compensated out-domain PLDA system is shown to perform as well as in-domain PLDA training with as few as 500 unlabelled in-domain i-vectors for NIST-2010 SRE and 2000 unlabelled in-domain i-vectors for NIST-2008 SRE, and considerable relative improvement over both out-domain and in-domain PLDA development if more are available.

Investigating in-domain data requirements for PLDA training

This paper analyzes the limitations upon the amount of in- domain (NIST SREs) data required for training a probabilistic linear discriminant analysis (PLDA) speaker verification system based on out-domain (Switchboard) total variability subspaces. By limiting the number of speakers, the number of sessions per speaker and the length of active speech per session available in the target domain for PLDA training, we investigated the relative effect of these three parameters on PLDA speaker verification performance in the NIST 2008 and NIST 2010 speaker recognition evaluation datasets. Experimental results indicate that while these parameters depend highly on each other, to beat out-domain PLDA training, more than 10 seconds of active speech should be available for at least 4 sessions/speaker for a minimum of 800 speakers. If further data is available, considerable improvement can be made over solely out-domain PLDA training.

Transcription of multi-genre media archives using out-of-domain data

We describe our work on developing a speech recognition system for multi-genre media archives. The high diversity of the data makes this a challenging recognition task, which may benefit from systems trained on a combination of in-domain and out-of-domain data. Working with tandem HMMs, we present Multi-level Adaptive Networks (MLAN), a novel technique for incorporating information from out-of-domain posterior features using deep neural networks. We show that it provides a substantial reduction in WER over other systems, with relative WER reductions of 15% over a PLP baseline, 9% over in-domain tandem features and 8% over the best out-of-domain tandem features. © 2012 IEEE.

Spatial domain mitigation of out of band strong interferers in HF wide band acquisition using analog beamforming principles

We envision that dynamic multiband transmissions taking advantage of the receiver diversity (even for collocated antennas with different polarization or radiation pattern) will create a new paradigm for these links guaranteeing high quality and reliability. However, there are many challenges to face regarding the use of broadband reception where several out of band (with respect to multiband transmission) strong interferers, but still within the acquisition band, may limit dramatically the expected performance. In this paper we address this problem introducing a specific capability of the communication system that is able to mitigate these interferences using analog beamforming principles. Indeed, Higher Order Crossing (HOCs) joint statistics of the Single Input ? Multiple Output (SIMO) system are shown to effectively determine the angle on arrival of the wavefront even operating over highly distorted signals.

Contribution of transmembrane domain V amino acids to β1l-adrenoceptor activity and affinity

Introduction. There are two binding sites on the β1-adrenoceptor (AR), β1H and β1L corresponding to high and low affinity binding sites respectively, which can be activated to cause cardiostimulation. Some β-blockers that block β1AR and β2ARs can activate β1LARs at higher concentrations than those required to cause blockade. The β2AR does not form a corresponding low affinity binding site and therefore we postulated that heterologous amino acids are responsible for the formation of β1LAR. Aim. To investigate whether heterologous amino acids of transmembrane domain V (TMDV) of β1AR and β2ARs contribute to β1LAR. Methods. β1ARs, β2ARs and mutant β1ARs containing all (β1(β2TMDV)AR) or single amino acids of TMDV of the β2AR were prepared and stably expressed in Chinese Hamster Ovary cells. Concentration-effect curves for cyclicAMP accumulation were carried out for (-)-CGP12177 in the absence or presence of (-)-bupranolol. Results. The potencies (pEC50) of (-)-CGP12177 were β2AR (9.24 ± 0.14, n = 5), β1(V230I)AR (9.07 ± 0.07, n = 10), β1(β2TMDV)AR (8.86 ± 0.10, n = 15), β1(R222Q)AR (8.09 ± 0.29, n = 6), β1AR (8.00 ± 0.11, n = 11). The affinities (pKB) of (-)-bupranolol were β2AR (9.82 ± 0.52, n = 5), β1(V230I)AR (7.64 ± 0.12, n = 8), β1(β2TMV)AR (8.06 ± 0.17, n = 8), β1(R222Q)AR (7.33 ± 0.23, n = 5), β1AR (7.23 ± 0.23, n = 5). Discussion. The potency of (-)-CGP12177 was higher at β2AR than at β1AR consistent with activation through a low affinity site at the β1AR (β1LAR). The presence of V230 in β1AR accounted for the lower potency of (-)-CGP 12177. The affinity of (-)-bupranolol was lower at β1AR compared to β2AR. The presence of V230 in β1AR accounted in part for the lower affinity. In conclusion TMDV of the β1AR contributes in part to the low affinity binding site of β1AR.

Contribution of transmembrane domain V amino acids to β1l-adrenoceptor activity and affinity

There are two binding sites on the β1-adrenoceptor (AR), β1H and β1L corresponding to high and low affinity binding sites respectively, which can be activated to cause cardiostimulation (reviewed Kaumann and Molenaar, 2008). Some β-blockers that block β1AR and β2ARs can activate β1LARs at higher concentrations than those required to cause blockade. The β2AR does not form a corresponding low affinity binding site (Baker et al 2002) and therefore we postulated that heterologous amino acids are responsible for the formation of β1LAR. Our aim was to investigate whether heterologous amino acids of transmembrane domain V (TMDV) of β1AR and β2ARs contribute to β1LAR. β1ARs, β2ARs and mutant β1ARs containing all (β1(β2TMDV)AR) or single amino acids of TMDV of the β2AR were prepared and stably expressed in Chinese Hamster Ovary cells. Concentration-effect curves for cyclicAMP accumulation were carried out for (-)-CGP12177 or (-)-isoprenaline in the absence or presence of (-)-bupranolol. _______________________________________________________________________ (-)-CGP 12177 (-)-Bupranolol affinity (pKB) pEC50 vs (-)-CGP 12177 vs (-)-isoprenaline _______________________________________________________________________ β1AR 8.00 ± 0.11 (11) 7.23 ± 0.23 (5) 9.52 ± 0.28 (5) β2AR (high density) 9.24 ± 0.14 (5) 9.82 ± 0.52 (8) xPaulxxxxxxx β2AR (low density) no effect β1(β2TMV)AR 8.86 ± 0.10 (15) 8.06 ± 0.17 (8) 9.08 ± 0.22 (6) β1(V230I)AR 9.07 ± 0.07 (10) 7.64 ± 0.12 (8) 9.36 ± 0.28 (9) β1(R222Q)AR 8.09 ± 0.29 (6) 7.33 ± 0.23 (5) 9.36 ± 0.08 (6) β1(V230A)AR 7.59 ± 0.09 (6) 7.32 ± 0.24 (4) 8.62 ± 0.18 (5) _______________________________________________________________________ The potency of (-)-CGP12177 was higher at β2AR than at β1AR consistent with activation through a low affinity site at the β1AR (β1LAR) but not β2AR. The presence of V230 in β1AR accounted for the lower potency of (-)-CGP 12177. The affinity of (-)-bupranolol at β1AR and mutants was higher when determined with (-)-isoprenaline than with (-)-CGP 12177. The affinity of (-)-bupranolol determined against (-)-CGP 12177 was lower at β1AR compared to β2AR. The presence of V230 in β1AR accounted in part for the lower affinity. In conclusion V230 of the β1AR contributes in part to the low affinity binding site of β1AR. Baker JG, Hall IP, Hill SJ (2002). Pharmacological characterization of CGP12177 at the human β2-adrenoceptor. Br J Pharmacol 137, 400−408 Kaumann AJ, Molenaar P (2008) The low-affinity site of the β1-adrenoceptor and its relevance to cardiovascular pharmacology. Pharmacol Ther 118, 303-336

Domain-specific application analysis for customized instruction identification

With the increasing importance of Application Domain Specific Processor (ADSP) design, a significant challenge is to identify special-purpose operations for implementation as a customized instruction. While many methodologies have been proposed for this purpose, they all work for a single algorithm chosen from the target application domain. Such algorithm-specific approaches are not suitable for designing instruction sets applicable to a whole family of related algorithms. For an entire range of related algorithms, this paper develops a methodology for identifying compound operations, as a basis for designing “domain-specific” Instruction Set Architectures (ISAs) that can efficiently run most of the algorithms in a given domain. Our methodology combines three different static analysis techniques to identify instruction sequences common to several related algorithms: identification of (non-branching) instruction sequences that occur commonly across the algorithms; identification of instruction sequences nested within iterative constructs that are thus executed frequently; and identification of commonly-occurring instruction sequences that span basic blocks. Choosing different combinations of these results enables us to design domain-specific special operations with different desired characteristics, such as performance or suitability as a library function. To demonstrate our approach, case studies are carried out for a family of thirteen string matching algorithms. Finally, the validity of our static analysis results is confirmed through independent dynamic analysis experiments and performance improvement measurements.

Doing something about it: Representations of NAPLAN in the public domain

Since 2008 all Australian school students have sat standardised tests in Reading, Writing, Language Conventions (Spelling, Grammar and Punctuation) and Numeracy in years 3,5,7 and 9. NAPLAN tests report individual students' attainment of skills against a set of standards. Individual student results are communicated to parents. Schools are then ranked against other schools depending upon the aggregate of their NAPLAN results. The process is explained to parents and community members as “improving the learning outcomes for all Australian students” (MCEETYA, 2009). This paper will examine NAPLAN as it is being played out in a mediated space through analysing unsolicited comment found in new media such as Twitter and online forums. NAPLAN intersects with contemporary debates about Australian education policy: the roles schools should play in improving national productivity, the relationship between state and federal government interest in education, the role and expectations of the teacher, what curriculum and pedagogy should be and look like and how limited financial resources can best be spread across education sectors and systems. These are not new considerations, however, what has changed is that education policy seems to have become even more of a political issue than it has before. This paper uses Ball's 'toolkit' approach to education policy analysis to suggest that there are multiple 'effects' of NAPLAN culminating in a series of disconnected conversations between various stakeholders.

Comparative kinomics of human and chimpanzee reveal unique kinship and functional diversity generated by new domain combinations

Background: Phosphorylation by protein kinases is a common event in many cellular processes. Further, many kinases perform specialized roles and are regulated by non-kinase domains tethered to kinase domain. Perturbation in the regulation of kinases leads to malignancy. We have identified and analysed putative protein kinases encoded in the genome of chimpanzee which is a close evolutionary relative of human. Result: The shared core biology between chimpanzee and human is characterized by many orthologous protein kinases which are involved in conserved pathways. Domain architectures specific to chimp/human kinases have been observed. Chimp kinases with unique domain architectures are characterized by deletion of one or more non-kinase domains in the human kinases. Interestingly, counterparts of some of the multi-domain human kinases in chimp are characterized by identical domain architectures but with kinase-like non-kinase domain. Remarkably, out of 587 chimpanzee kinases no human orthologue with greater than 95% sequence identity could be identified for 160 kinases. Variations in chimpanzee kinases compared to human kinases are brought about also by differences in functions of domains tethered to the catalytic kinase domain. For example, the heterodimer forming PB1 domain related to the fold of ubiquitin/Ras-binding domain is seen uniquely tethered to PKC-like chimpanzee kinase. Conclusion: Though the chimpanzee and human are evolutionary very close, there are chimpanzee kinases with no close counterpart in the human suggesting differences in their functions. This analysis provides a direction for experimental analysis of human and chimpanzee protein kinases in order to enhance our understanding on their specific biological roles.

Thermodynamic characterization of the reversible, two-state unfolding of maltose binding protein, a large two-domain protein

The folding and stability of maltose binding protein (MBP) have been investigated as a function of pH and temperature by intrinsic tryptophan fluorescence, far- and near-UV circular dichroism, and high-sensitivity differential scanning calorimetric measurements. MBP is a monomeric, two-domain protein containing 370 amino acids. The protein is stable in the pH range of 4-10.5 at 25 degrees C. The protein exhibits reversible, two-state, thermal and guanidine hydrochloride-mediated denaturation at neutral pH. The thermostability of MBP is maximal at pH 6, with a Tm of 64.9 degrees C and a deltaHm of 259.7 kcal mol(-1). The linear dependence of deltaHm on Tm was used to estimate a value of deltaCp of 7.9 kcal mol(-1) K(-1) or 21.3 cal (mol of residue)(-1) K(-1). These values are higher than the corresponding deltaCp's for most globular proteins studied to date. However, the extrapolated values of deltaH and deltaS (per mole of residue) at 110 degrees C are similar to those of other globular proteins. These data have been used to show that the temperature at which a protein undergoes cold denaturation depends primarily on the deltaCp (per mol of residue) and that this temperature increases with an increase in deltaCp. The predicted decrease in stability of MBP at low temperatures was experimentally confirmed by carrying out denaturant-mediated unfolding studies at neutral pH at 2 and 28 degrees C.

Performance Analysis of Reconstruction Techniques for Frequency-Domain Optical-Coherence Tomography

We address the issue of noise robustness of reconstruction techniques for frequency-domain optical-coherence tomography (FDOCT). We consider three reconstruction techniques: Fourier, iterative phase recovery, and cepstral techniques. We characterize the reconstructions in terms of their statistical bias and variance and obtain approximate analytical expressions under the assumption of small noise. We also perform Monte Carlo analyses and show that the experimental results are in agreement with the theoretical predictions. It turns out that the iterative and cepstral techniques yield reconstructions with a smaller bias than the Fourier method. The three techniques, however, have identical variance profiles, and their consistency increases linearly as a function of the signal-to-noise ratio.

Probing the aging effect in metallic polypyrrole by terahertz time-domain spectroscopy

Terahertz time-domain spectroscopy has been carried out on a metallic film of polypyrrole (PPy doped by PF6). The sample was exposed to air to investigate how the conductivity of the film varies as a function of time. The absorption and dispersion of the film decrease during initial days, and then tend to saturate. The conductivity of unaged sample follows the Drude model, and upon aging the data fit to the localization-modified Drude model. The fitting parameters show that the number of charge carriers decreases during the aging process. The initial rapid decrease in conductivity suggests that some of the delocalized carriers are localized due to aging. (C) 2007 American Institute of Physics.

Terahertz Time Domain Spectroscopy to Detect Low-Frequency Vibrations of Double-Walled Carbon Nanotubes

We have measured the frequency-dependent real index of refraction and extinction coefficient (and hence the complex dielectric function) of a free-standing double-walled carbon nanotube film of thickness 200 nm by using terahertz time domain spectroscopy in the frequency range 0.1 to 2.5 THz. The real index of refraction and extinction coefficient have very high values of approximately 52 and 35, respectively, at 0.1 THz, which decrease at higher frequencies. Two low-frequency phonon modes of the carbon nanotubes at 0.45 and 0.75 THz were clearly observed for the first time in the real and imaginary parts of the complex dielectric function along with a broad resonance centred at around 1.45 THz, the latter being similar to that in single-walled carbon nanotubes assigned to electronic excitations. Our experiments bring out a possible application of double-walled carbon nanotube films as a neutral density filter in the THz range.

Time-domain-finite-wave analysis of the engine exhaust system by means of the stationary-frame method of characteristics. Part I. Theory

Time-domain-finite-wave analysis of the engine exhaust system is usually done using the method of characteristics. This makes use of either the moving frame method, or the stationary frame method. The stationary frame method is more convenient than its counterpart inasmuch as it avoids the tedium of graphical computations. In this paper (part I), the stationary-frame computational scheme along with the boundary conditions has been implemented. The analysis of a uniform tube, cavity-pipe junction including the engine and the radiation ends, and also the simple area discontinuities has been presented. The analysis has been done accounting for wall friction and heat-transfer for a one-dimensional unsteady flow. In the process, a few inconsistencies in the formulations reported in the literature have been pointed out and corrected. In the accompanying paper (part II) results obtained from the simulation are shown to be in good agreement with the experimental observations.