This paper proposes a stereo echo canceller with simultaneous input-sliding and sliding-period control. Periodic delays with different phases in two channels provide an additional condition for echo cancellation, leading to faster convergence. It also guarantees symmetrical shifts of a sound image around its original position for better subjective quality of the far-end signal. Sliding-period control with more slides for less interchannel correlation of the input signals achieves less noticeable sliding artifacts and faster convergence simultaneously. Subjective evaluation results demonstrate that simultaneous sliding achieved comparable far-end signal (FES) quality to that of the original FES. It is also shown that sliding-period control tops up 14% improvement in convergence time without trading off the subjective quality of the FES.

Source separation and signal distortion are theoretically analyzed for the FF-BSS systems implemented in both the time and frequency domains and the FB-BSS system. The FF-BSS systems have some degree of freedom, and cause some signal distortion. The FB-BSS has a unique solution for complete separation and distortion free. Next, the condition for complete separation and distortion free is derived for the FF-BSS systems. This condition is applied to the learning algorithms. Computer simulations by using speech signals and stationary colored signals are carried out for the conventional methods and the new learning algorithms employing the proposed distortion free constraint. The proposed method can drastically suppress signal distortion, while maintaining high separation performance. The FB-BSS system also demonstrates good performances. The FF-BSS systems and the FB-BSS system are compared based on the transmission time difference in the mixing process. Location of the signal sources and the sensors are rather limited in the FB-BSS system.

Source separation and signal distortion in three kinds of BSSs with convolutive mixture are analyzed. They include a feedforward BSS, trained in the time domain and in the frequency domain, and a feedback BSS, trained in the time domain. First, an evaluation measure of signal distortion is discussed. Second, conditions for source separation and distortion free are derived. Based on these conditions, source separation and signal distortion are analyzed. The feedforward BSS has some degree of freedom, and the output spectrum can be changed. The feedforward BSS, trained in the frequency domain, has weighting effect, which can suppress signal distortion. This weighting is, however, effective only when the source spectra are similar to each other. Since, the feedforward BSS, trained in the time domain, does not have any constraints on signal distortion free, its output signals can he easily distorted. A new learning algorithm with a distortion free constraint is proposed. On the other hand, the feedback BSS can satisfy both source separation and distortion free conditions simultaneously. Simulation results support the theoretical analysis.

This paper proposes a stereophonic acoustic echo canceller without pre-processing which can identify the correct echo-paths. By dividing the filter coefficients into two portions and updating one part at a time, the filter coefficient have a unique solution. Convergence analysis clarifies the condition for correct echo-path identification. For fast convergence and stable adaptation, a convergence detection and an adaptive step-size are also introduced. The modification amount of the filter coefficients detects the convergence and also determines the step-size. Computer simulations show 10dB smaller coefficient error than those of the conventional algorithms.

A network structure and its learning algorithm have been proposed for blind source separation applied to nonlinear mixtures. Nonlinearity is expressed by low-order polynomials, which are acceptable in many practical applications. A separation block and a linearization block are cascaded. In the separation block, the cross terms are suppressed, and the signal sources are separated in each group, which include its high-order components. The high-order components are further suppressed through the linearization block. A learning algorithm minimizing the mutual information is applied to the separation block. A new learning algorithm is proposed for the linearization block. Simulation results, using 2-channel speech signals, instantaneous mixtures, and 2nd-order post nonlinear functions, show good separation performance.

A learning process of a single neural network (SNN) to improve prediction accuracy of protein secondary structure is optimized. The protein secondary structures are predicted using a multiple alignment of amino acid as the input data. A multi-modal neural network (MNN) has been proposed to improve the precision of prediction. This method uses five independent neural networks, and the final decision is made by averaging all outputs of five SNNs. In the proposed method, the same prediction accuracy can be achieved by using only a single NN and optimizing a learning process. In a learning process of protein structure prediction, over learning is easily occurred. So, the learning process is optimized so as to avoid the over learning. For this purpose, small learning rates, adding small random noise to the input data, and updating the connection weights by the average in some group are useful. The prediction accuracy 58% obtained by using the conventional SNN is improved to 66%, which is the same accuracy of the MNN, which needs five SNNs.

In order to achieve fast convergence and less computation for adaptive filters, a joint method combining a whitening process and the NLMS algorithm is a hopeful approach. However, updating the filter coefficients is not synchronized with the reflection coefficient updating, resulting in unstable behavior. We analyze the effects of this, and propose the "synchronized learning algorithm" to solve this problem. The synchronous error between them is removed, and fast convergence and small residual error are obtained. This algorithm, however, requires O(ML) computations, where M is an adaptive filter length, and L is a lattice predictor length. It is still large compared with the NLMS algorithm. In order to achieve less computation while the fast convergence is maintained, a block implementation method is proposed. The reflection coefficients are updated at some period, and are fixed during this interval. The proposed block implementation can be effectively applied to parallel form adaptive filters, such as sub-band adaptive filters. Simulation using speech signals shows that the learning curve of the proposed block implementation is a little slower than our original algorithm, but can save computational complexity.

First, convergence properties in blind source separation (BSS) of convolutive mixtures are analyzed. A fully recurrent network is taken into account. Convergence is highly dependent on relation among signal source power, transmission gain and delay in a mixing process. Especially, reverberation degrade separation performance. Second, a learning algorithm is proposed for this situation. In an unmixing block, feedback paths have an FIR filter. The filter coefficients are updated through the gradient algorithm starting from zero initial guess. The correction is exponentially scaled along the tap number. In other words, stepsize is exponentially weighted. Since the filter coefficients with a long delay are easily affected by the reverberations, their correction is suppressed. Exponential weighting is automatically adjusted by approximating an envelop of the filter coefficients in a learning process. Through simulation, good separation of performance, which is the same as in no reverberations condition, can be achieved by the proposed method.

A network structure and its learning algorithm have been proposed for blind source separation applied to nonlinear mixtures. The network has a cascade form consists of a source separation block and a linearization block in this order. The conventional learning algorithm is employed for the separation block. A new learning algorithm is proposed for the linearization block assuming 2nd-order nonlinearity. After, source separation, the outputs include the nonlinear components for the same signal source. This nonlinearity is suppressed through the linearization block. Parameters in this block are iteratively adjusted based on a process of solving a 2nd-order equation of a single variable. Simulation results, using 2-channel speech signals and an instantaneous nonlinear mixing process, show good separation performance.

In order to achieve fast convergence and less computation for adaptive filters, a joint method combining a whitening process and the NLMS algorithm is a hopeful approach. One way is to combine a lattice predictor and a transversal filter supervised by the NLMS algorithm. However, the filter coefficient adaptation is very sensitive to the reflection coefficient fluctuation. The reason for this instability is analyzed. The filter coefficients are updated one sample behind the reflection coefficient update. This causes large error, in other words, sensitivity of their mismatch is very high on filter characteristics. An improved learning method is proposed in order to compensate for this mismatch. The convergence property is close to that of the RLS algorithm. Computational complexity can be well reduced from that of the RLS algorithm. Simulation results using real voices demonstrate the usefulness of the proposed method.

Many problems solved by multilayer neural networks (MLNNs) are reduced into pattern mapping. If the mapping includes several different rules, it is difficult to solve these problems by using a single MLNN with linear connection weights and continuous activation functions. In this paper, a structure trainable neural network is proposed. The gate units are embedded, which can be trained together with the connection weights. Pattern mapping problems, which include several different mapping rules, can be realized using a single new network. Since some parts of the network can be commonly used for different mapping rules, the network size can be reduced compared with the modular neural networks, which consists of several independent expert networks.

In Jutten and Herault's (1991) blind separation algorithm, symmetrical distribution and statistical independence of the signal sources are assumed. When they are not satisfied, the learning process becomes unstable. In order to avoid the unstable behavior, two stabilization methods are proposed. Since large samples easily disturb symmetrical distribution, the outputs of the separation process with large amplitude are detected, and the learning is skipped. Imbalance of the signal source powers affects statistical independence. It is estimated by the cross-correlation of the observed signals. When the cross-correlation is high, the correction term by the above algorithm algorithm becomes wrong. Therefore, adjusting the weights in the separation process is skipped. Computer simulation using many kinds of signal sources demonstrates the signal sources with asymmetrical distribution and imbalanced power are well separated.

This paper presents convergence characteristics of stereophonic echo cancellers with pre-processing. The convergence analysis of the averaged tap-weights show that the convergence characteristics depends on the relation between the impulse response in the far-end room and the changes of the pre-processing filters. Examining the uniqueness of the solution in the frequency domain leads us to the same relation. Computer simulation results show the validity of these analyses.

In this paper, a polyphase and FFT realization of a sub-band adaptive filter is proposed. In the polyphase and FFT realization for the transmultiplexer, the sampling rate reduction is the same as the number of the sub-bands. In the proposed method, however, they are different, so the conventional method cannot be applied. In the new realization, the polyphase filters are divided into a tapped delay line and an FFT part so as to be shared by all the sub-bands. Computational complexity is well reduced compared with the direct realization.

Network size of neural networks is highly dependent on activation functions. A trainable activation function is proposed, which consists of a linear combination of some basic functions. The activation functions and the connection weights are simultaneously trained. An 8-bit parity problem can be solved by using a single output unit and no hidden unit. In this paper, we expand this model to multilayer neural networks. Furthermore, nonlinear functions are used at the unit inputs in order to realize more flexible transfer functions. The previous activation functions and the new nonlinear functions are also simultaneously trained. More complex pattern classification problems can be solved with a small number of units and fast convergence.

A robust adaptive microphone array (RAMA) using a new adaptation-mode control method (AMC) is proposed, and its evaluation results by hardware are presented. The adaptation of the RAMA is controlled based on an SNR (signal-to-noise) estimate using the output powers of the fixed beamformer and the adaptive blocking matrix. The RAMA is implemented on a multi-DSP real time signal-processing system with a C-compiler. Simulation results with real acoustic data show that the AMC based on the SNR estimate causes less breathing noise than the conventional AMC and that it obtains 1.0-point higher score on a 5-point mean opinion score scale. Evaluation through a real time signal-processing system demonstrates that noise reduction achieved by the RAMA is over 12 dB even in reverberant environments.

This paper presents a new robust adaptive microphone array (AMA) and its evaluation in an echoic environment. The proposed AMA is a generalized sidelobe canceller equipped with a variable blocking matrix using coefficient-constrained adaptive filters, and a multiple-input canceller using norm-constrained adaptive filters (NCAFs). Because the NCAFs have selective nonlinearity in the relationship between coefficient norm and coefficient error, the proposed AMA has better spatial selectivity than the conventional AMA. Evaluation with real acoustic data captured in a room of 0.3-second reverberation time shows that the noise is suppressed by 19 dB. In subjective evaluation, the proposed AMA obtains 3.8 on a 5-point mean opinion score scale.

This paper presents the implementation and performance evaluation of sparse-tap adaptive FIR filters with tap-position control. For the reduction of computational costs, a small number of filter coefficients are assigned to significant regions in the impulse response. A floating-point digital signal processor (DSP) ADSP-21020 is used for the implementation. For one second of end-to-end delay and 32 milliseconds of total significant-region length, the number of computations is reduced to almost one-tenth compared with a full-tap FIR adaptive filter using the same DSP. Experimental results using a telephone network simulator show that the implemented echo canceller reduces the echo by almost 20 dB for both white-noise and real speech signals. Tracking performance for phase rolls and flat-delay changes is also examined.

An accurate method recovering the 3D object function g(X,Y,D) from an image sequence of continuous focus series is proposed. Not only the depth map D(x,y) but also a pan-focused brightness map g/sub p/(X,Y) (a clear texture image observed by an ideal pinhole camera) are obtained. The depth map (2 1/2-D sketch) of the scene is evaluated by detecting that a distinct maximum of the point spread function exists in the focal plane. To estimate the maximum, a local variance of the brightness distribution is computed as a function of the focus location. A clear texture of the scene (object) is composed by collecting the brightness values from the respective focused 128 (or 256) frames. Reconstruction of the 3-D sketch is tested by mapping the texture on the depth map. The usefulness of a polarized filter is also demonstrated to recover an accurate depth map in the presence of specular reflection.

This paper proposes a new noise-robust adaptive FIR filtering algorithm with an adaptive step-size which takes non-stationarity of speech signals into account. The proposed algorithm controls the step-size based on the reference input signal power and the estimated noise power. Implementation of an acoustic echo canceller based on this algorithm using digital signal processors is also described. Computer simulation results using a real speech signal show that the proposed algorithm improves the ERLE (echo return loss enhancement) by more than 10 dB compared with conventional noise-robust adaptive-step algorithms. The implemented echo canceller achieves 25 dB of the ERLE, which agrees with the computer simulation results.

This paper presents a stereo echo canceller implemented by digital signal processors (DSPs) and its performance evaluation. The implemented echo canceller, which takes the cross-correlation between stereo signals into account, has a single adaptive filter per channel. This structure reduces the number of adaptive filters by half compared with the conventional stereo echo canceller which requires two adaptive filters per channel. Analytical results with real impulse responses clarify the conditions in which this echo canceller achieves better performance with a smaller number of taps than the conventional one. Applicable arrangements of microphones and a talker at the far-end are shown. Evaluation by the implemented hardware supports the analytical results. Satisfactory performance is obtained for both white-noise and real-speech inputs.>

A Hi-Fi audio codec with an improved adaptive transform coding (ATC) algorithm is presented using digital signal processors (DSPs). An audio signal with a 20 kHz bandwidth sampled at 48 kHz is coded at a rate of 128 kb/s. The algorithm utilizes adaptive block size selection, which is effective for preecho suppression. A modified discrete cosine transform (MDCT) with a simple window set is employed to reduce block boundary noise without decreasing the performance of transform coding. In addition, a fast MDCT calculation algorithm, based on a fast Fourier transform, is adopted. Weighted bit allocation is employed to quantize the transformed coefficients. The codec was realized by a multiprocessor system composed of newly developed DSP boards. Subjective tests with the codec show that the coding quality is comparable to that of compact disc signals.>