A continuous function model is developed for efficient head-related impulse response (HRIR) representation. In this model, HRIR is modeled as the response of an IIR filter via common factor decomposition analysis, where HRIRs at the same azimuth but different elevations share the same IIR pole part and HRIRs at the same elevation but different azimuthes share the same IIR zero part. The pole and zero coefficients of the proposed common factor IIR (CF-IIR) filters are further modeled using sinusoidal functions of azimuth and elevation, respectively. A fast algorithm for evaluating sinusoidal function for efficient synthesis of virtual 3D sound is also developed.

In this paper, we present a novel post-processing scheme to improve autoregressive (AR) speech parameter estimation in noise. The proposed technique exploits temporal correlation between dynamic all-pole parameters to capture natural speech evolution. To achieve this, a Kalman tracking scheme is proposed to track line spectrum frequency (LSF) trajectories with system parameter learned directly from processed online data. To facilitate the online system identification, a heuristic approach is initially proposed to preliminarily remove “musical tones” caused by conventional frame-based methods in LSF domain. Through performance evaluation based on a study of spectrogram and objective measures, it is demonstrated that the proposed post-processing scheme successfully restores natural and smooth evolution of speech dynamics, and in the meantime, effectively removes processing artifacts caused by conventional methods in various conditions.

In this letter, a Two-Dimension Common Factor Decomposition (2D-CFD) algorithm is proposed to represent the 3D (time, azimuth and elevation) head-related impulse response (HRIR) dataset with two factorized matrices. One matrix is made up of a set of common factor responses extracted from HRIRs with the same azimuth angle and carries azimuthal information contained in original HRIR dataset. The other matrix carries elevation information similarly. By using this dimension reduction, the storage requirement of HRIRs is reduced remarkably. In addition, the common factor responses in the matrix are further modeled with efficient low order IIR filters to reduce the computation complexity in 3D sound synthesis.

In this paper, we propose two block-based clustering and identification algorithms that contribute to robust estimation of autoregressive (AR) speech parameters in noisy environments. Motivated by the fact that the evolution pattern of speech dynamics could be an observable feature that are retained in a series of noisy observations, a dynamic state tracking scheme based on Kalman filter is incorporated to utilize this additional trajectory information in block-based AR codebook design. The proposed algorithm is devised in a sense that AR blocks with similar clean line spectrum frequency trajectories as well as noisy-to-clean mappings are clustered offline and identified online. It is compared with conventional vector quantization based approaches that directly minimize a distortion between AR parameters. Through objective assessments based on mean square error and log-spectral distance, it is demonstrated that the proposed algorithm achieves significant improvement over conventional methods in various conditions.

In this work, we present a model-based approach to enhance noisy speech using an analysis-synthesis framework. Target speech is reconstructed with model parameters estimated from noisy observations. In particular, spectral envelope is estimated by tracking its temporal trajectories in order to improve the noise-distorted short-time spectral amplitude. Initially, we propose an analysis-synthesis framework for speech enhancement based on harmonic noise model (HNM). Acoustic parameters such as pitch, spectral envelope, and spectral gain are extracted from HNM analysis. Spectral envelope estimation is improved by tracking its line spectrum frequency trajectories through Kalman filtering. System identification of Kalman filter is achieved via a combined design of codebook mapping scheme and maximum-likelihood estimator with parallel training data. Complete system design and experimental validations are given in details. Through performance evaluation based on a study of spectrogram, objective measures and a subjective listening test, it is demonstrated that the proposed approach achieves significant improvement over conventional methods in various conditions. A distinct advantage of the proposed method is that it successfully tackles the “musical tones” problem.

This paper presents a Kalman tracking approach to re-estimate clean spectral amplitude from noisy speech spectrum for re-synthesis based speech enhancement. The motivation of using Kalman filter and training is to exploit the temporal correlation between speech dynamics and to include prior knowledge of speech to improve the model parameter estimation in harmonic noise model (HNM) based speech enhancement system. The re-estimated harmonic amplitude is fitted into an analysis-synthesis framework to accomplish a more accurate HNM based re-synthesis. Objective evaluation results show the proposed method achieves significant improvement over various classical short-time spectral amplitude (STSA) based methods, especially in low signal-to-noise ratio environments.

In this paper, Principal Component Analysis (PCA) is performed on HRIRs grouped according to azimuth and covariance to improve distortion performance while reducing storage requirement. By further combining Balanced Model Truncation (BMT) using a joint optimization scheme, a virtual 3D sound synthesis scheme with low computation complexity, little storage requirement and efficient linear interpolation is proposed. A smaller model distortion is achieved as high PCA model error of contralateral HRIR is decreased by grouping PCA strategy.

In this paper, frequency estimation of a two-dimensional (2D) cisoid in the presence of additive white Gaussian noise is addressed. By utilizing the rank-one property of the 2D noise-free data matrix, the frequencies are estimated in a separable manner from the principal left and right singular vectors according to an iterative weighted least squares procedure. We have also derived the mean and variance expressions for the frequency estimates, which show that they are approximately unbiased and their accuracy achieves Cramér-Rao lower bound (CRLB) at sufficiently high signal-to-noise ratio conditions. Computer simulation results are included to corroborate the theoretical development as well as to contrast the performance of the proposed algorithm with the weighted phase averager and iterative quadratic maximum likelihood method as well as CRLB.

The problem of frequency estimation for noisy sinusoidal signals from multiple segments or channels, which are referred to as gapped data, is addressed. Based on linear prediction and weighted least squares techniques, an iterative relaxation-based frequency estimator is devised and analyzed. The proposed algorithm is also extended to harmonically related frequencies. Computer simulations are conducted to compare the estimation performance of the developed approach with an existing multichannel frequency estimator and Crame¿r-Rao lower bound.

This paper presents a speech enhancement method based on an analysis-synthesis framework using harmonic noise model (HNM) in car noise environment. The major advantages of this method are effective suppression of car noise even in very low signal-to-noise ratio environments and mitigation of ldquomusical tonesrdquo which are generally introduced by conventional methods. In this paper, we devise a complete analysis-synthesis based speech enhancement system, and give details in HNM modeling, parameter estimation, and car noise adaptation. Subjective evaluation results show that the proposed method exhibits better noise suppression ability over conventional approaches without obvious degradation of speech quality.

Extending narrowband speech (0-4 kHz) to wideband speech (0-8 kHz) has applications in telephone systems and speech recognition systems where wideband training speech data may not be available. A couple of methods have been proposed to retrieve the missing high-band information (4-8 kHz) from narrowband speech. Memoryless systems are likely to produce large hissing artifacts since mutual information between low-band (0-4 kHz) and high-band (4-8 kHz) spectra are actually quite low. Generally speaking, bandwidth extension cannot recover original high-band information but good approximates with less over-estimation of the high-band energy, which usually refers to hissing artifact, can be obtained by considering the neighboring speech frames. In this paper, we propose a new bandwidth extension system with memory by using a state-space model to capture the long-term speech dynamics. The model parameters can be trained in the sense of maximum likelihood (ML) and the enhancement is obtained via wideband state vector estimation and Kalman filtering. The performance in terms of spectral distortion is shown to be much better than other memoryless systems and is comparable with early Continuous Density Hidden Markov Model (CDHMM) memory system. The new state-space method is inherent sequential and has advantages of less processing delays and robustness against block detection errors.

In this paper we present a block-based bandwidth extension system to enhance the quality of narrowband speech signal (0-4 kHz). In memoryless bandwidth extension systems, the missing high-band components are estimated from narrowband speech using the current frame only. As the narrowband-to-wideband mapping is a one-to-many problem, this memoryless system is likely to cause hissing and whistling artifacts in the reproduced speech. Our method estimates high-band components via narrowband-to-wideband state sequence mapping using a continuous density hidden Markov model (CDHMM) on a block basis. The speech block is either one word or a sequence of words in a narrowband utterance. CDHMM estimation method avoids the one-to-many property of low-band and high-band dependency. Both subjective and objective evaluations show that hissing and whistling artifacts are reduced and the spectrally extended wideband speech (0-8 kHz) is pleasant to listen.

This paper presents a procedure to improve the quality of narrowband (0–4khz) CELP coded speech. The procedure is based to refine the pitch periodicity and reinsert the high frequency components (4–8khz) in the narrowband CELP decoded speech. The narrowband CELP decoded speech is first analyzed with Harmonic+Noise analyzer and Lowband information are extracted. By exploiting the Lowband spectrum envelope and V/UV information, the highband (4–8khz) spectrum envelope is recovered statistically by using a voiced/unvoiced gaussian mixture model with interpolation. Lowband information along with the estimated highband information is then fed to the Hannonic+Noise synthesizer to re-synthesize a wideband speech. The objective and subjective tests are performed to evaluate the quality of the re-synthesis wideband (0–8khz) speech. The results of the above experiments show that the re-synthesis wideband speech is pleasant to listen with crispy characteristics and preferred over the CELP coded speech.

A high quality harmonic+noise speech coder is presented. This coder operates in either the steady mode or the transient mode. An efficient phase coding technique based on polynomial fitting is proposed. In this technique, the frequency deviations that are used to minimize the phase prediction errors are modeled by a 6/sup th/-order polynomial. The polynomial coefficients are quantized in closed-loop by a split vector quantizer. For the transient speech, the unvoiced part is modeled by an 8/sup th/-order all-pole model while the voiced part is modeled by the low-frequency harmonic components. The transitional behavior is characterized by the onset time and growth rate of each harmonic component. An analysis-by-synthesis procedure is proposed to determine the onset times and growth rates in the time domain. The proposed coder retains the temporal events of the speech signal and produces high quality speech at 2.4 kbps.

This paper presents a harmonic+noise speech coder which uses an efficient spectral quantization technique and a novel voiced/unvoiced (V/UV) mixing model. The harmonic magnitudes are coded at 23 bits/frame using the magnitude response of a linear predictive coding (LPC) system. The difference between the harmonic magnitudes and the sampled magnitude response is minimized by the closed-loop approach. The V/UV mixing is modeled by a smooth function which is derived from the speech spectrum envelope based on the flatness measure. The V/UV mixing model allows noise to be added in the harmonic portion of speech spectrum so that buzzyness is reduced. The V/UV mixing information is determined from the spectral parameters available in the decoder, no bits are needed for transmitting the V/UV information. A 1.4 kbps harmonic coder is developed. The speech quality of the coder is comparable to other harmonic coders operating at higher rates.

Results for improving the quality of narrowband CELP-coded speech by enhancing the pitch periodicity and by regenerating the high-band components of speech spectra are reported. Multiband excitation (MBE) analysis is applied to enhance the pitch periodicity by re-synthesizing the speech signal using a harmonic synthesizer. The high-band magnitude spectra are regenerated by matching to low-band spectra using a trained wideband spectral codebook. Information about the voiced/unvoiced (V/UV) excitation in the high-band is derived from a training procedure and recovered by using the matched low-band index. Simulation results indicate that the quality of the wideband enhanced speech is significantly improved over the narrowband CELP-coded speech.

A variable bit rate multiband excited linear predictive speech coder is proposed in this paper. Speech signals are compressed at different bit rates ranging from 0.88 kbps to 2.6 kbps according to the mode of operation and the optimum V/UV transition frequency. An average bit rate of 1.24 kbps is achieved. The proposed speech coder improves the speech quality by splitting the non-stationary speech segments for analysis. The V/UV distribution of a short-time speech spectrum is represented efficiently by using a closed-loop minimised V/UV transition frequency. Depending on the V/UV transition frequency, the spectrum envelope is quantized at variable bit rates through embedded differential predictive scalar and vector quantizations of the LSP parameters. The proposed spectral quantization scheme results in a spectral distortion comparable to a fixed 24-bit 2-dimensional differential scalar quantization scheme.

In this paper, a high-quality wideband speech coder is proposed. The coding structure resembles a LD-CELP coder, however, several novel improvements are made. The gain adapter for the stochastic codebook is driven by a neural network and it updates the excitation gain in a sample-by-sample fashion. The purpose of incorporating a neural network is to exploit both the intraand inter-frame correlation of speech signal in a non-linear manner. A psychoacoustic model instead of a simple perceptual weighting filter is used to shape the quantization noise. Simulation result shows that the proposed coder can achieve transparent coding of wideband speech at 16 kbps.

This paper describes a hybrid input/output block adaptation scheme for the quantization of spectral parameters for low-delay CELP coders. In this scheme, line spectrum pair (LSP) coefficients are used as parameters for spectral adaptation and they are partitioned into two segments; LSPs in the first segment which corresponding to low frequency regions are backward adapted from the reproduction speech while LSPs in the second segment are adapted from input speech and quantized as side information to the decoder. The rationale behind this scheme is to compensate for the spectrum distortion introduced by conventional backward adaptation schemes operating at low bit rates. Simulation results show that the hybrid LSP adaptation scheme achieves transparent quantization at 12 bits for each analysis frame of 5 ms.

A high quality speech synthesis method based on interpolating the voiced/unvoiced (V/UV) mixture functions of the multiband excitation model (MBE) is proposed. In the MBE model, each harmonic band of the fundamental frequency in an excitation spectrum is rigidly declared as either voiced or unvoiced and the harmonic band is pitch-dependent. In the proposed method, each harmonic band in a short time spectrum is synthesized by mixing both voiced and unvoiced energies. The ratio of the V/UV energies in a spectrum is determined by the V/UV mixture function which is subsequently parametrized by an all-zero model. Since the V/UV decision in the proposed method is not rigidly declared and the V/UV mixture function is pitch-independent, interpolating the V/UV excitation spectrum becomes possible. Smooth transition of excitation between acoustic units can be achieved by interpolating the V/UV mixture functions of adjacent frames. Simulation results show that by incorporating the V/UV mixture function for concatenative synthesis, significant improvement in synthetic speech quality can be achieved.

A fast codebook search method for code excited linear predictive (CELP) coding of speech is described. The method relies on using a vector-sum codebook where the crosscorrelation of any pair of basis vectors is odd-symmetric. Due to this odd-symmetric crosscorrelation (OSC) property the energy term of the cost function for codebook search is a constant with respect to the search, and a simple sign detection procedure is used to locate the optimum codeword with a complexity almost independent of the codebook size. An algorithm for generating a set of OSC basis vectors is described. Simulation results show that by replacing the standard VSELP codebooks with the OSC codebooks, the same coder performance can be achieved but with a much lower complexity. An OSC-CELP coder was implemented and demonstrated to achieve good quality speech at rates below 4.8 kbps.

For low bit rate linear predictive speech coding, it is desirable to quantize the LPC parameters using as few bits as possible with transparent quantization quality. A novel structural constrained vector quantization scheme, namely split residual vector quantization (SRVQ), is proposed. In this scheme, PARCOR coefficient vector is split into subvectors and each subvector is sequentially quantized according to a multi-stage structure based on the principle of minimizing the prediction error energies. Simulation results show that a 2-stage split residual vector quantizer can achieve transparent quantization at 25 bit/frame while autocorrelation analysis is employed to calculate the LPC parameters.>

A novel split vector quantization (SVQ) scheme for low bit rate coding of speech signals is proposed. In this scheme, the LPC parameter vector, which is represented by Parcor coefficients, is split into small-dimension subvectors, and each subvector is sequentially quantized according to a multistage structure that resembles a segmented lattice filter. The forward and backward prediction residuals in the segmented filter are coupled across VQ stages. The quantizer in each stage operates on the principle of minimizing the forward and backward prediction error energies similar to linear predictive analysis. Simulation results show that the new split VQ scheme can achieve transparent quantization of LPC parameters at 25 b/frame.>

The authors introduce a spectral analysis technique for LPC systems based on using a high-order lattice filter having only few nonzero reflection coefficients. An efficient algorithm based on dynamic programming techniques has been developed to find the nonzero coefficients and their corresponding delays for the thinned lattice filter. Simulation results show that the thinned filter approach can achieve more accurate spectrum fitting to speech signals.>