This paper reports on the optimal design of the excitation codebook in the 13 kbps variable rate QCELP coder of Korean speech. We present two optimal excitation codebooks which consist of 128 and 556 samples, respectively. For the design and test of the improved codebook, a data base of Korean speech is used. A quasi-Newton optimization algorithm was developed to design the codebook. The optimized codebook which remains sparse, can produce an average gain of 0.84 and 0.45 dB in SNR and SEGSNR respectively. Informal listening tests confirm the improvement in speech quality.

This paper addresses traffic prediction issues on MPEG. A new adaptive traffic prediction scheme is proposed using MPEG picture characteristic that picture traffic depends on the coding mode of that picture, that is, I, P, and B mode. Our prediction scheme, which is based n picture decomposition (PD) and the cross-correlation of the different types of pictures, has better performance in predicting bursty MPEG traffic than that of the first-order autoregressive (AR) prediction scheme. Our simulation results show that the performance is further improved about 15% by utilizing the cross-correlations between pictures.

One fo the major problems in speech recognition is the mismatch between training and testing environments. Recently, SNR normalization technique, which normalizes the dynamic range of frequency channels in mel-scaled filterbank, was proposed[1]. While it showed improved robustness against additive noise, it requires a reliable speech detection mechanism and several adaptation parameters to be optimized. In this paper, we propose a modified SNR normalization technique. In this technique, we take simply the maximum of filterbank output and predetermined masking constant for each frequency band. According to the speaker-independent isolated word recognition in car noise environments, proposed modification yields better recognition performance that the original SNR normalization method, with rather reduced complexity.

The goal of postprocessing in automatic speech recognition is to improve recognition performance by utterance verification at the output of recognition stage. It is focused on the effective rejection of out-of vocabulary words based on the confidence score of hypothesized candidate word. We present two methods for computing confidence scores. Both methods are based on the distance between each observation vector and the representative code vector, which is defined by the most likely code vector at each state. While the first method employs simple time normalization, the second one uses a normalization technique based on the concept of on-line garbage mode[1]. According to the speaker independent isolated words recognition experiment with discrete density HMM, the second method outperforms both the first one and conventional likelihood ratio scoring method[2].

A mobile propagation characteristics for wideband DS-SS (Direct Sequence-Spectrum) signal is presented. Existing narrowband model is extended for the wideband pulse with an arbitrary shape. The received DS-SS signal in the frequency domain is the transfer function of the propagation channel weighted by the inverse Fourier transform. In this proposed method, received signal spectral density, instantaneous waveform, and Doppler spectrum of DS-SS signal via either Rayleigh of Rician channel can be obtained easily. Simulation results match well with both simulated theoretical fading statistics and classical theory. As expected, the extraction of chip timing in Rician fading shown to be more tractable than Rayleigh fading.

This study models mobile communication channels as a discrete finite Markovian process, and Markovian jump linear system having parallel Kalman filter type is applied. What is newly proposed in this paper is an equation for obtaining the transition matrix according to sampling time by using a weighted Gaussian sum approximation and its simple calculation process. Experiments show that the proposed method has superior performance and reuires computation compared to the existing MJLS using the ransition matrix given by a statistical method or from priori information.

A frequency-variable ultrasonic vibrator is designed, which is made of a multi-layered PZT vibrator with a conical horn. Transmission line equations for the conical horn are derived and analyzed using the equivalent circuit method in order to analyze the characteristics of the vibrator. The controllability of the driving characteristics by varying the electrical impedance is confirmed by the experimental results of the free admittance characteristics and the vibrational velocity distributions.

In this paper, we propose an adaptive encoding method of fixed codebook in CELP coders and implement an adaptive fixed code exited linear prediction(AF-CELP) speech coder. AF-CELP exploits the fact that the fixed codebook contribution to speech signal is also periodic like the adaptive codebook (or pitch filter) contribution. By modeling the fixed code book with the pitch lag and the gain from the adaptive codebook, AF-CELP can be implemented at low bit rates as well as low complexity. Listening tests show that a 6.4 kbit/s AF-CELP has a comparable quality to the 8 kbit/s CS-ACELP in background noise conditions.

Based on the anthropometric data of Korea male adults, a head and torso simulator(HATS) is constructed to measure its head related transfer functions (HRTFs) which can be used for three dimensional (3-D) sound localization. The HRTFs binaural impulse responses, are measured in an anechoic chamber using a burst maximum length sequence (MLS) signal of 65,535 samples and 32,768 samples acquisition at the sampling rate of 75.47kHz. Also measured are the impulse responses of a driving loudspeaker and some headphones for sound reproduction to get the exact HRTF of the HATS-alone. Through a post-processing procedure, the impulse-version HRTFs at the sampling frequency of 44.1 kHz, which have filter lengths of 512 points, are finally obtained. As an application of the measured HRTFs, a 3-D sound processor for headphone reproduction has been developed. The signal intervals to be processed can be selected and each interval is manipulated to have its diretionality and distance information by using corresponding HRTF and energy control.

In this paper, we describe a real-time implementation of a multi-channel auido codec system that is based on the MPEG-1 audio algorithm. The major feature of this system is that it has a flexible multi-DSP system that can be adapted for various applications with using up to four TMS320C40 DSPs. The purpose of this paper is to present the problems of the system and is to describe the optimized methods to solve the problems in the view of hardware and software. Our audio codec is composed of an encoder an a decoder system and the bit rate of bitstream is up to 384 kbps. Fast input/output interfaces, DSP overloads, and inter-DSP communications methods with high speed are considered in multi-DSP H/W. Also, to run real-time in S/W, optimizing methods of algorithm are considered. After implementation of system, the subjective assessment method, and 'triple stimulus/hidden reference/double blind' that recommended by ITU-R TG10/3 is adopted for the quality of our system. All test items except one are awarded difference grades(diffgrade) better than 1-. Form the results, multi-channel audio system can be used for HDTV service.

In this paper, a modified adaptive sign (MAS) algorithm based on a mixed norm error criterion is proposed. The mixed norm error criterion of be minimized is constructed as a combined convex function of the mean-absolute error and the mean-absolute error to the third power. A convergence analysis of the MAS algorithm is also presented. Under a set of mild assumptions, a set of nonlinear evolution equations that characterizes the statistical mean and mean-squared behavior of the algorithm is derived. Computed simulations are carried out to verify the validity of our derivations.