Sensitivity weighted vector quantization of line spectral pair frequencies

In previous work, a method to compensate the divergence between the distributions of natural and generated modulation spectra (MS) has been proposed for hidden Markov model (HMM) based speech synthesis. This method can alleviate the over-smoothing effect of parameter generation when mel-cepstral coefficients (MCC) are used as spectral features. This paper further investigates the MS compensation method for line spectral pairs (LSP). Four approaches to extract MS from LSPs are implemented and compared. These approaches calculate MS vectors using original LSP sequences, log power spectra (LPS) derived from LSPs, MCCs derived from LSPs, and MCCs derived from speech waveforms, respectively. Experimental results show that the naturalness of synthetic speech gets improved after MS compensation when LSPs are used as spectral features for HMM modeling. The degree of improvement depends on the type of spectral features for MS calculation significantly. MCCs derived from LSPs are more suitable for MS compensation than original LSPs and LPS derived from LSPs. Besides, using MCCs derived from speech waveforms also achieves satisfactory performance. This means that MS compensation can also be implemented as a post-filter to synthetic waveforms which does not rely on the type of spectral features and vocoders adopted in the synthesis system.

This paper describes a novel algorithm for transforming linear prediction coefficients (LPCs) to line spectral frequencies (LSFs) and line spectral pairs (LSPs) used by most of the speech processing applications. The symmetric and antisymmetric polynomials (SAPS) for LSP/LSFs, corresponding to the LPC polynomial, are first multiplexed into a single real sequence. The required samples of SAPS correspond to the DFT of the obtained real sequence. The proposed algorithm is referred as PMLS as it is based on the Plus Minus (PM) algorithm an FFT which efficiently computes the DFT of a real sequence for the positive frequency interval only. The samples of the SAPS are efficiently utilized for the computation of a single parameter which is used for computation of LSF and LSP independently with some interpolation principles. This interpolation exploits the available samples of SAPS and does not require their samples at finer resolution. The efficiency of the PMLS is illustrated with the help of some examples. Some guidelines for an optimal implementation of PMLS on fixed point digital signal processors (DSPs) are also presented.

I invented a non-negative temporal decomposition method for line spectral pairs and articulatory parameters based on the multiplicative update rules. These parameters are decomposed into a set of temporally overlapped unimodal event functions restricted to the range [0,1] and corresponding event vectors. When line spectral pairs are used, event vectors preserve their ordering property. With the proposed method, the RMS error of the measured and reconstructed articulatory parameters is 0.21 mm and the spectral distance of the measured and reconstructed line spectral pairs parameters is 2.0 dB. The RMS error and spectral distance in the proposed method are smaller than those in conventional methods. This technique will be useful for many applications of speech coding and speech modification.

An elegant relationship (here called symmetry) between the linear predictive coding (LPC) analysis and the composite sinusoidal modeling (CSM) analysis is described. The objective is to provide a systematic view of the theory of voice spectral modeling such as LPC, PARCOR, CSM, and line spectral pairs (LSP). All of these methods of analysis are reduced to the problems of deriving the model parameters for lower‐order autocorrelation functions equal to the number of degrees of freedom of the model. From the point of view of the theory of orthogonal polynomials, LPC and CSM are the theory of orthogonal polynomials of a unit circle (in the case of LPC) and on the real axis (in the case of CSM). Beautiful symmetries hold in regard to the formulation, definition of various parameters, and solution algorithm. Also, a new interpretation is given here for the LSP from the point of view of orthogonal polynomials. © 2002 Scripta Technica, Electron Comm Jpn Pt 3, 85(6): 42–54, 2002

A fundamental issue in real-time interactive voice transmissions over unreliable IF networks is the loss or late arrival of packets for playback. This problem is especially serious when transmitting low bit rate-coded speech with pervasive dependencies introduced. In this case, the loss or late arrival of a single packet will lead to the loss of subsequent dependent frames. We study end-to-end loss-concealment schemes for ensuring high quality in playback. We propose a novel multiple description-coding method for concealing packet losses in transmitting low bit rate-coded speech. Based on high correlations observed in linear predictor parameters-in the form of Line Spectral Pairs (LSPs)-of adjacent frames, we generate multiple descriptions in senders by interleaving LSPs, and reconstruct lost LSPs in receivers by linear interpolations. As excitation codewords have low correlations, we further enlarge the segment size for excitation generation and replicate excitation codewords in all the descriptions in order to maintain the same transmission bandwidth. Our proposed scheme can be extended easily to more than two descriptions and can adapt its number of descriptions dynamically to network-loss conditions. Experimental results on FS-1016 CELP, ITU G.723.1, and FS MELP coders show good performance of our scheme.

Speech signal is the continuous signal which has the characteristics in its own. For recognizing the speech signal, the system must be able to classify and recognize the speech feature. Almost of this system is speaker-independent speech system. This paper presents statistical methods for speech recognition by extracting the features of the speech and analyzing their trajectory. The speech feature has been extracted to line spectral pairs (LSP) coefficients and then uses the statistic model to pattern the trajectory for recognizing the signal. The result shows that the using technique usually works well which the maximum accuracy of recognition is 99.67% at number 1 of male speech and the minimum accuracy of recognition is 82.33% at number 5 of female speech.

This paper presents a class of quadratically weighted distortion measures which provide optimal performance for the high rate vector quantization (VQ) of linear predictive coding (LPC) parameters. It is shown that the quantization distortion of a high rate VQ converges to a quadratically weighted measure, where the quadratic weighting matrix is a matrix, which is a generalization of the scalar sensitivity concept to the vector case. The sensitivity matrix is the second order term of the Taylor series expansion of the original distortion measure. Closed form expressions and computationally efficient methods for computing the sensitivity matrices of the different LPC parameterizations are given, which involve no numerical integration and can be implemented in real-time on modern DSP chips. In the general case, the sum of sensitivity weighted scalar errors is not equivalent to the original distortion measure. However the sensitivity matrix of the line spectral pair (LSP) frequencies is exactly diagonal, demonstrating that for LSPs only a sum of sensitivity weighted scalar errors will result in optimal performance.

To improve the performance of the HMM-based voice conversion system in which the LSP coefficient is introduced as the spectral representation, a model clustering technique to tie HMMs into classes for the model adaptation, considering the phonetic and linguistic contextual factors of HMMs, is adopted in this paper. Besides, due to the relationship between the LSP coefficients of adjacent orders, an appropriate format of the regression matrix is suggested according to the small amount of the adaptation training data. Subjective and objective tests prove that the source HMMs can be adapted more accurately using the proposed method, meanwhile the synthetic speech generated from the adapted model has better discrimination and speech quality. Index Terms : model adaptation, regression matrix clustering, and regression matrix format 1. Introduction With the development of the corpus-based speech synthesis technique, the intelligibility and naturalness of the synthetic speech has been improved a lot. However, it is still a difficult problem for the corpus-based TTS system to synthesize speech of various speakers and speaking styles with a limited database. So the voice conversion technique which can convert one speaker’s voice to another speaker’s voice provides a positive approach to achieve the goal of synthesizing speech of multi-speakers. The HMM-based voice conversion system is built on the basis of the HMM-based speech synthesis. In the HMM-based speech synthesis system, spectrum, pitch and duration are modeled simultaneously in a unified framework of HMMs [1][2][3]. In addition, voice characteristics of the synthetic speech can be converted from one speaker to another by applying a model adaptation algorithm, such as the MLLR (maximum likelihood linear regression) algorithm [4][5], with a small amount of speech uttered by the target speaker. We have realized a HMM-based speech synthesis system in which the LSP (line spectral pair) coefficients and the STRAIGHT (Speech Transformation and Representation using Adaptive Interpolation of weighted spectral contour) analysis-synthesis algorithm are introduced [6][7]. Then, by realizing the MLLR algorithm, we provide our synthesis system with the ability of synthesizing voice of various speakers. However, there still exist two main problems in the HMM-based voice conversion system. Firstly, the data-driven clustering method described in the MLLR algorithm ignores many contextual factors between HMMs, therefore some unrelated HMMs are forced into one class which will affect the accuracy of the model adaptation. Secondly, the system performance including the voice characteristics and voice quality of the synthetic speech decreases greatly when the adaptation training data is very limited. In order to solve these problems, a clustering method, considering the phonetic and linguistic connections between HMMs using the context decision tree, which has been applied similarly in both the HMM-based speech recognition and the HMM-based speech synthesis areas [8][9], is described in this paper. Moreover, an appropriate regression matrix format is suggested when very few training data is available, as the LSP coefficients of only several adjacent orders have strong correlations. In the following part of this paper, an overview of our HMM-based voice conversion system is presented in section 2. Section 3 describes the details of the proposed context clustering decision tree and the appropriate regression matrix for the model adaptation. Section 4 presents the results of experiments including subjective and objective evaluations while section 5 provides a final conclusion.

We present a study of a channel optimized matrix quantizer (COMQ) applied to quantize LSP (line spectral pair) parameters in a CELP (code excited linear prediction) coder. A modification of the DoD FS-1016 standard (Campbell et al., 1989) is used for this purpose. A Gaussian channel is considered as the channel through which information is sent and BPSK (binary phase shift-keying) is used as modulation technique. To measure the performance of this coder several channel signal to noise ratios are considered. Also, the performance of the FS-1016 standard coder is computed for comparison purposes.

Postfilters for suppressing coding noises in low bit rate speech coders usually operate in the time domain with high complexity and also introduce nonlinear phase distortion to the speech signals because they are recursive filters. In this paper, a frequency-domain postfiltering technique for multiband excited linear predictive (MBELP) coders is proposed. The idea is to firstly sample the postfilter magnitude spectrum at the pitch harmonics and then scale the signal band magnitudes by the sampled postfilter magnitudes. This approach does not need to specifically perform the filtering process, and the signal energies before and after postfiltering can be easily equalized by simply re-scaling the signal band magnitudes. Another obvious advantage of using this frequency-domain postfiltering technique is that the phase information utilized in multiband synthesis will be retained because only the magnitude spectra are modified. In this paper, a fast method is given to sample the postfilter magnitude spectrum which is characterized by line spectral pair (LSP) parameters. The proposed frequency-domain postfilter has a much lower implementation complexity than its time-domain counterpart.

Many coding techniques have been proposed for the compression of speech signals. Speech signals are continuous signals shaped by the movement of the vocal tract. Linear predictive coefficients (LPC) and line spectral pairs (LSP) coefficients have been used to model the position of the vocal tract. Each speaker's vocal tract uses a unique, finite set of positions to generate their voice. This paper compares of the performance of ordered and disordered codebooks which employed to measure and classify the repetition of the speech signal coefficients. An analysis and results of the efficient performance of both types of codebooks in terms of the quality of synthesised speech are also provided. Also both ordered and disordered codebooks can reduce the number of bit rate transmission around 20% in speech coding system

Accurate modeling of co-articulation, the contextsensitive merging of the boundaries between allophones in continuous speech, is vital for natural sounding speech synthesis. This paper describes initial research investigating the use of Bezier Curves to form models of co-articulation in human speech. A 12th order, pitch synchronous line spectral pair (LSP) [1] analysis is performed on a corpus of 239 phonetically balanced sentences of English speech. The resulting data are divided to form an inventory of the diphones occurring in the speech database. The trajectory of each line spectral pair parameter through each diphone can then be represented by a single cubic Bezier curve segment, found using the LevenbergMarquardt curve fitting method [2, 3]. Results are presented showing the accuracy of Bezier models of the coarticulation between different types of speech sounds.

Artificial neural networks (ANNs), in combination with parametric spectral representation techniques, are applied for the detection of helicopter sound. Training of the ANN detectors was based on simulated helicopter sound from four helicopters and a variety of nonhelicopter sounds. Coding techniques based on linear prediction coefficients (LPCs) have been applied to obtain spectral estimates of the acoustic signals. Other forms of the LPC parameters such as reflection coefficients, cepstrum coefficients, and line spectral pairs (LSPs) have also been used as feature vectors for the training and testing of the ANN detectors. We have also investigated the use of wavelet transform for signal de-noising prior to feature extraction. The performance of various feature extraction techniques is evaluated in terms of their detection accuracy.

The Internet has become a vital source of information. This comes with the attendant problems, namely misuse. Systematic downloading of academic and digitized media have become commonplace. Academic institutions in particular get blacklisted owing to free availability of Internet across campus. The objective of this paper is to pro-actively detect systematic downloading. Time series of number of requests are analyzed with pattern analysis techniques. A characterization of the model in the Z-domain shows that the roots of the transfer function form separate clusters during normal and abnormal behavior of traffic. Stability of the system has been used as a cue to detect systematic downloading. Analyzing the trajectory and location of roots to detect systematic downloading involves complex decisions which may not be robust to evolving traffic. This issue is addressed by using Support Vector Data Description that learns a hypersphere enclosing normal traffic. Our empirical evaluations with features like AutoRegressive model roots and Line Spectral Pairs (LSP) obtained from the request time series along with data description show a lot of promise for detecting systematic downloading with F-measure and accuracy as high as 0.90 and 99.5%, respectively. This hybrid approach ensures low false alarms, misses and guarantees robustness of the system.

The use of line spectral pair (LSP)-frequencies for approaching the performance of a full vector quantizer for speech coding is discussed. An algorithm is presented that attains close to optimal performance if an accumulative constraint is introduced in the parameter encoding. Numerical results are presented from the training of code-books as well as from the evaluation of the quantizer. Some listening tests are discussed. >