The acoustic characteristics of acoustic tubes with protrusions at the radiation end are computed by FEM simulation. In the rst experiment, two di erent shapes of the protrusions, a symmetrical and an asymmetrical shape with respect to the vertical, are investigated. Frequency characteristics of the radiation impedance are computed from simulation results. The simulation results show that the results of FEM simulation are in good agreement with our measurement results. The proposed 3-D radiational model is useful for analysis of the acoustic characteristics of human speech. In the 2nd experiment, the protrusion is attached to our 3-D vocal tract model. The vocal tract shape corresponds to the Japanese vowel /a/. The cross sections of the tubes are eliptic in shape. The simulation results show that the vocal tract transfer function of the FEM results is di erent from our previous FEM results and 1-D analytical solution.

This article addresses two critical problems in designing new full-duplex fast training algorithms for simultaneously estimating echo and channel responses. Although algorithms of this type have been discussed and evaluated in previous work, the proposed algorithms concentrate on coping with the following two problems that were not solved in the previous approaches: (1) the symbol rate difference between the local transmitter and the remote transmitter, and (2) the frequency offsets in both far echo and far signal that are caused by the analog carrier network. The performance of the new methods is analyzed in terms of the mean square error. Simulation results are presented to confirm the analysis.

The problem of reducing the initialization time of the echo cancellers and the channel equalizers in modems is extremely important for high-speed full-duplex data transmission over two-wire lines. In a previous publication Chen et al. (see IEEE Trans. Information Theory, vol.41, no.5, 1995) proposed a full-duplex fast training procedure for simultaneously estimating the echo and channel responses. It reduces the tap-setting time to half of that required by the traditional half-duplex fast training schemes. However, its estimation accuracy may be degraded by the frequency offsets in both the far echo and far signal that are caused by the analog carrier network. We extend the previous work and develop a new algorithm that can compensate for the frequency offsets in both the far echo and far signal. The performance of the new method is analyzed in terms of the mean-square error. Simulation results are presented to confirm the analysis.

To reduce the system initialization time in high-speed full-duplex data transmission over two-wire lines, we propose an efficient full-duplex fast training algorithm which can simultaneously estimate the impulse responses of echo paths and channels at both ends. Two mutually orthogonal periodic sequences are designed and used to co-estimate the near echo, the far echo with bulk delay, and the channel response. The new algorithm can reduce the tap-setting time to half of that required by the traditional half-duplex fast training schemes. The effects of channel noise and symbol rate offset between two ends are examined in terms of both mean-square error (MSE) and signal-to-noise ratio (SNR). Both the theoretical analysis and computer simulation show that there are three degrading factors due to symbol rate offset. The SNR is inversely proportional to the sum of the estimated coefficients, the half period of the training sequences, and the square of symbol rate offset. If the far-signal-to-channel-noise ratio is 30 dB, then the degradation is significant when the symbol rate offset is more than 10/sup -4/. If it is 40 dB, then the degradation is significant when the symbol rate offset is above 3/spl times/10/sup -5/.>

An efficient algorithm for maximum likelihood (ML) bearing estimation of narrowband signal sources using a uniform linear array of sensors is proposed. This algorithm adopts the quasi-Newton method which is one of the most efficient gradient methods. The estimates are obtained by computing zeros of a polynomial whose coefficients can be given by solving the nonlinear optimization problem arising from the ML criterion. Simulations have indicated the following properties: the proposed method gives better estimation accuracy than MUSIC, gives almost the same accuracy as the conventional ML procedure alternating projection (AP), and requires lower computational cost than AP.>

Finite-element analysis of the vibration of vocal cords is considered. The air flow from the lung to the vocal tract across the glottis is assumed to be a viscous fluid flow, and the muscles of the vocal cords are assumed to be viscoelastic bodies. By considering the viscosity and nonlinearity in the fluid system, the adaptable loss of the air-flow appears. A method of total simulation of the fluid and structural systems is proposed, and, as an example, one simplified problem is simulated. The complicated deformed shapes of the vocal cord model are obtained.>

In the vocal tract model, the area function of vocal tracts is accurately represented by a polynomial function with only a few parameters, and the digital filter realization includes the loss, the accurate lip impedance and the variation of the tract length. We measured the lip impedance using a manikin and pipes set in a baffle, and obtained the reflection coefficient at the lips. From the results of these experiments, we discuss the difference between the measured characteristics and those of a spherical baffle model, and show a method to obtain approximate digital filters for the reflection characteristics at the end of the vocal-tract; i.e. the lips. By introducing this filter, it becomes possible to express arbitrary changes of the length of the vocal-tract without changing the sampling frequency.

A vocal-tract model with boundary conditions is proposed, and it is shown that the model can be applied to a vocal-tract simulator add an estimation algorithm of vocal-tract shapes. Since our vocal-tract simulator can dynamically synthesize continuous speech with the vocal-fold vibration and the loss factor in the vocal-tract, this simulator can be used as a tool for accurate evaluation of the estimation accuracy of the formants or the characteristics of the vocal-tract. We propose a new method estimating cross-sectional area of vocal-tract from the ARMA parameters, and this method is based on the vocal-tract model and our ARMA estimation algorithms.

This paper introduces a new ARMA digital lattice filter. It can be designed from the recursive formulae which solve estimation error filter equations for an ARMA model. This filter consists of two elementary sections. These sections are an AR type elementary section which realizes the AR part of the estimated ARMA model, and an MA type elementary section which realizes the MA part of the estimated ARMA model. By using these sections, we can design the ARMA digital lattice filter with an arbitrary AR and an arbitrary MA order, and easily test its stability. In addition, some experimental results for a real speech waveform are shown.

Recently, a model identification system ( MIS ) has been studied for ARMA parameter estimation. Since in MIS an estimator with a high-order ARMA parameter and a model reduction algorithm are used, the tracking ability of the spectral estimation is good. However in real speech analysis, estimation errors are caused by the use of a low-pass filter for A/D conversion. In order to compensate for such errors we propose a weighted MIS ( WMIS ) which includes a compensation for the characteristics of the low-pass filter. The WMIS estimates the input of the reference model and time-varying ARMA parameters, and achieves the property of rapid convergence by using a high-order model. Furthermore the algorithm for minimum realization is shown as a model reduction algorithm. The proposed algorithms are applied to synthetic speech and real speech, and it is shown that the estimated spectra sufficiently represent the variation of formants without jitters in the high-frequency part.