Unsupervised training for mandarin broadcast news and conversation transcription

A significant cost in obtaining acoustic training data is the generation of accurate transcriptions. For some sources close-caption data is available. This allows the use of lightly-supervised training techniques. However, for some sources and languages close-caption is not available. In these cases unsupervised training techniques must be used. This paper examines the use of unsupervised techniques for discriminative training. In unsupervised training automatic transcriptions from a recognition system are used for training. As these transcriptions may be errorful data selection may be useful. Two forms of selection are described, one to remove non-target language shows, the other to remove segments with low confidence. Experiments were carried out on a Mandarin transcriptions task. Two types of test data were considered, Broadcast News (BN) and Broadcast Conversations (BC). Results show that the gains from unsupervised discriminative training are highly dependent on the accuracy of the automatic transcriptions. © 2007 IEEE.

Development of low entropy coding in a recurrent network.

In this paper we present an unsupervised neural network which exhibits competition between units via inhibitory feedback. The operation is such as to minimize reconstruction error, both for individual patterns, and over the entire training set. A key difference from networks which perform principal components analysis, or one of its variants, is the ability to converge to non-orthogonal weight values. We discuss the network's operation in relation to the twin goals of maximizing information transfer and minimizing code entropy, and show how the assignment of prior probabilities to network outputs can help to reduce entropy. We present results from two binary coding problems, and from experiments with image coding.

The effect of wake passing on a flow separation in a low pressure turbine cascade

This paper describes an investigation into the effect that passing wakes have on a separation bubble that exists on the pressure surface and near the leading edge of a low pressure turbine blade. Previous experimental studies have shown that the behaviour of this separation is strongly incidence dependent and that it responds to its disturbance environment. The results presented in this paper examine the effect of wake passing in greater detail. Two dimensional, Reynolds averaged, numerical predictions are first used to examine qualitatively the unsteady interaction between the wakes and the separation bubble. The separation is predicted to consist of spanwise vortices whose development is in phase with the wake passing. However, comparison with experiments shows that the numerical predictions exaggerate the coherence of these vortices and also overpredict the time-averaged length of the separation. Nonetheless, experiments strongly suggest that the predicted phase locking of the vortices in the separation onto the wake passing is physical.

The effects of a trip wire and unsteadiness on a high speed highly loaded low-pressure turbine blade

This paper presents the effect of a single spanwise 2D wire upon the downstream position of boundary layer transition under steady and unsteady inflow conditions. The study is carried out on a high turning, high-speed, low pressure turbine (LPT) profile designed to take account of the unsteady flow conditions. The experiments were carried out in a transonic cascade wind tunnel to which a rotating bar system had been added. The range of Reynolds and Mach numbers studied includes realistic LPT engine conditions and extends up to the transonic regime. Losses are measured to quantify the influence of the roughness with and without wake passing. Time resolved measurements such as hot wire boundary layer surveys and surface unsteady pressure are used to explain the state of the boundary layer. The results suggest that the effect of roughness on boundary layer transition is a stability governed phenomena, even at high Mach numbers. The combination of the effect of the roughness elements with the inviscid Kelvin-Helmholtz instability responsible for the rolling up of the separated shear layer (Stieger [1]) is also examined. Wake traverses using pneumatic probes downstream of the cascade reveal that the use of roughness elements reduces the profile losses up to exit Mach numbers of 0.8. This occurs with both steady and unsteady inflow conditions.

Development of blade profiles for low pressure turbine applications

This paper describes a program of work, largely experimental, which was undertaken with the objective of developing an improved blade profile for the low-pressure turbine in aero-engine applications. Preliminary experiments were conducted using a novel technique. An existing cascade of datum blades was modified to enable the pressure distribution on the suction surface of one of the blades to be altered. Various means, such as shaped inserts, an adjustable flap at the trailing edge, and changing stagger were employed to change the geometry of the passage. These experiments provided boundary layer and lift data for a wide range of suction surface pressure distributions. The data was then used as a guide for the development of new blade profiles. The new blade profiles were then investigated in a low-speed cascade that included a set of moving bars upstream of the cascade of blades to simulate the effect of the incoming wakes from the previous blade row in a multistage turbine environment.

Unsteady shock motion on a NACA0012 aerofoil at low reduced frequencies

Using transonic blowdown windtunnel experiments, the 2D unsteady shock motion on a NACA0012 aerofoil is examined at various frequencies typical for helicopter blades in forward flight. The aerofoil is subjected to freestream velocities oscillating periodically between M = 0.66 and M = 0.77. Unsteady pressure traces and schlieren images are analyzed over a range of low reduced frequencies to provide information on shock location and strength throughout the cycle. Unsteady effects were noticeable even at very low reduced frequencies (down to O(0.01). However, through the range of frequencies investigated, and within experimental error, the unsteady shock location showed no discernible lag compared to the quasi-steady behaviour. On the other hand, significant variations were observed in shock strengths with the upstream running part of the cycle (decreasing Mach number) displaying considerably stronger shocks than during the accelerating part of the cycle. It could be shown that this variation in shock strength is primarily caused by the shock motion modifying the relative shock Mach number. As a result is was possible to use the quasi-steady results to predict the unsteady shock behaviour at the frequencies investigated here (below 0(0.1)).

Unsteady shock motion on a NACA0012 aerofoil at low reduced frequencies

Using transonic blowdown windtunnel experiments, the 2D unsteady shock motion on a NACA0012 aerofoil is examined at various frequencies typical for helicopter blades in forward flight. The aerofoil is subjected to freestream velocities oscillating periodically between M = 0.66 and M = 0.77. Unsteady pressure traces and schlieren images are analyzed over a range of low reduced frequencies to provide information on shock location and strength throughout the cycle. Unsteady effects were noticeable even at very low reduced frequencies (down to O(0.01). However, through the range of frequencies investigated, and within experimental error, the unsteady shock location showed no discernible lag compared to the quasi-steady behaviour. On the other hand, significant variations were observed in shock strengths with the upstream running part of the cycle (decreasing Mach number) displaying considerably stronger shocks than during the accelerating part of the cycle. It could be shown that this variation in shock strength is primarily caused by the shock motion modifying the relative shock Mach number. As a result is was possible to use the quasi-steady results to predict the unsteady shock behaviour at the frequencies investigated here (below 0(0.1)).

The low Reynolds number aerodynamics of leading edge flaps

Within the low Reynolds number regime at which birds and small air vehicles operate (Re=15,000-500,000), flow is beset with laminar separation bubbles and bubble burst which can lead to loss of lift and early onset of stall. Recent video footage of an eagle's wings in flight reveals an inconspicuous wing feature: the sudden deployment of a row of feathers from the lower surface of the wing to create a leading edge flap. An understanding of the aerodynamic function of this flap has been developed through a series of low speed wind tunnel tests performed on an Eppler E423 aerofoil. Experiments took place at Reynolds numbers ranging from 40000 to 140000 and angles of attack up to 30°. In the lower range of tested Reynolds numbers, application of the flap was found to substantially enhance aerofoil performance by augmenting the lift and limiting the drag at certain incidences. The leading edge flap was determined to act as a transition device at low Reynolds numbers, preventing the formation of separation bubbles and consequently decreasing the speed at which stall occurs during landing and manoeuvring.

Turbulent flow structure in experimental laboratory wind-generated gravity waves

This paper is the third part of a report on systematic measurements and analyses of wind-generated water waves in a laboratory environment. The results of the measurements of the turbulent flow on the water side are presented here, the details of which include the turbulence structure, the correlation functions, and the length and velocity scales. It shows that the mean turbulent velocity profiles are logarithmic, and the flows are hydraulically rough. The friction velocity in the water boundary layer is an order of magnitude smaller than that in the wind boundary layer. The level of turbulence is enhanced immediately beneath the water surface due to micro-breaking, which reflects that the Reynolds shear stress is of the order u *w 2. The vertical velocities of the turbulence are related to the relevant velocity scale at the still-water level. The autocorrelation function in the vertical direction shows features of typical anisotropic turbulence comprising a large range of wavelengths. The ratio between the microscale and macroscale can be expressed as λ/Λ=a Re Λ n, with the exponent n slightly different from -1/2, which is the value when turbulence production and dissipation are in balance. On the basis of the wavelength and turbulent velocity, the free-surface flows in the present experiments fall into the wavy free-surface flow regime. The integral turbulent scale on the water side alone underestimates the degree of disturbance at the free surface. © 2012 Elsevier B.V.

Investigation of low nanosecond light-matter interaction mechanisms

Laser ablation of solid Silicon targets using pulsed Yb fiber lasers of pulse duration 1.5-400 ns Yb fiber lasers is studied in this work. Material responses of a range of pulse envelopes are examined including front peak (FP) and double peak (DP) pulses. Theoretical models for the interactions are examined and qualitative explanations of material response experiments are presented.

Blade loading and its application in the mean-line design of low pressure turbines

In order to minimize the number of iterations to a turbine design, reasonable choices of the key parameters must be made at the earliest possible opportunity. The choice of blade loading is of particular concern in the low pressure (LP) turbine of civil aero engines, where the use of high-lift blades is widespread. This paper presents an analytical mean-line design study for a repeating-stage, axial-flow Low Pressure (LP) turbine. The problem of how to measure blade loading is first addressed. The analysis demonstrates that the Zweifel coefficient [1] is not a reasonable gauge of blade loading because it inherently depends on the flow angles. A more appropriate coefficient based on blade circulation is proposed. Without a large set of turbine test data it is not possible to directly evaluate the accuracy of a particular loss correlation. The analysis therefore focuses on the efficiency trends with respect to flow coefficient, stage loading, lift coefficient and Reynolds number. Of the various loss correlations examined, those based on Ainley and Mathieson ([2], [3], [4]) do not produce realistic trends. The profile loss model of Coull and Hodson [5] and the secondary loss models of Craig and Cox [6] and Traupel [7] gave the most reasonable results. The analysis suggests that designs with the highest flow turning are the least sensitive to increases in blade loading. The increase in Reynolds number lapse with loading is also captured, achieving reasonable agreement with experiments. Copyright © 2011 by ASME.

Low-order modelling of ducted flames with temporally varying equivalence ratio in realistic geometries

The interaction between unsteady heat release and acoustic pressure oscillations in gas turbines results in self-excited combustion oscillations which can potentially be strong enough to cause significant structural damage to the combustor. Correctly predicting the interaction of these processes, and anticipating the onset of these oscillations can be difficult. In recent years much research effort has focused on the response of premixed flames to velocity and equivalence ratio perturbations. In this paper, we develop a flame model based on the socalled G-Equation, which captures the kinematic evolution of the flame surfaces, under the assumptions of axisymmetry, and ignoring vorticity and compressibility. This builds on previous work by Dowling [1], Schuller et al. [2], Cho & Lieuwen [3], among many others, and extends the model to a realistic geometry, with two intersecting flame surfaces within a non-uniform velocity field. The inputs to the model are the free-stream velocity perturbations, and the associated equivalence ratio perturbations. The model also proposes a time-delay calculation wherein the time delay for the fuel convection varies both spatially and temporally. The flame response from this model was compared with experiments conducted by Balachandran [4, 5], and found to show promising agreement with experimental forced case. To address the primary industrial interest of predicting self-excited limit cycles, the model has then been linked with an acoustic network model to simulate the closed-loop interaction between the combustion and acoustic processes. This has been done both linearly and nonlinearly. The nonlinear analysis is achieved by applying a describing function analysis in the frequency domain to predict the limit cycle, and also through a time domain simulation. In the latter case, the acoustic field is assumed to remain linear, with the nonlinearity in the response of the combustion to flow and equivalence ratio perturbations. A transfer function from unsteady heat release to unsteady pressure is obtained from a linear acoustic network model, and the corresponding Green function is used to provide the input to the flame model as it evolves in the time domain. The predicted unstable frequency and limit cycle are in good agreement with experiment, demonstrating the potential of this approach to predict instabilities, and as a test bench for developing control strategies. Copyright © 2011 by ASME.