In studies on artificial bandwidth extension (ABE), there is a lack of international coordination in subjective tests between multiple methods and languages. Here we present the design of absolute category rating listening tests evaluating 12 ABE variants of six approaches in multiple languages, namely in American English, Chinese, German, and Korean. Since the number of ABE variants caused a higher-than-recommended length of the listening test, ABE variants were distributed into two separate listening tests per language. The paper focuses on the listening test design, which aimed at merging the subjective scores of both tests and thus allows for a joint analysis of all ABE variants under test at once. A language-dependent analysis, evaluating ABE variants in the context of the underlying coded narrowband speech condition showed statistical significant improvement in English, German, and Korean for some ABE solutions.

There is a considerable performance gap between the current scalable audio coding schemes and a nonscalable coder operating at the same bitrate. This suboptimality results from the independent coding of the layers in these systems. One of the aspects that plays a role in this suboptimality is the entropy coding. In practical audio coding systems including MPEG advanced audio coding (AAC), the transform domain coefficients are quantized using an entropy-constrained quantizer. In MPEG-4 scalable AAC (S-AAC), the quantization and coding are performed separately at each layer. In case of Huffman coding, the redundancy introduced by the entropy coding at each layer is larger at lower quantization resolutions. Also, the redundancy for the overall coder becomes larger as the number of layers increases. In fact, there is a tradeoff between the overall redundancy and the fine-grain scalability in which the bitrate per layer is smaller and more layers are required. In this paper, a fine-grain scalable coder for audio signals is proposed where the entropy coding of a quantizer is made scalable via joint design of entropy coding and quantization. By constructing a Huffman-like coding tree where the internal nodes can be mapped to the reconstruction points, the tree can be pruned at any internal node to control the rate-distortion (RD) performance of the encoder in a fine-grain manner. A set of metrics and a trellis-based approach is proposed to create a coding tree so that an appropriate path is generated on the RD plane. The results show the proposed method outperforms the scalable audio coding performed based on reconstruction error quantization as used in practical systems, e.g., in S-AAC.

Considering the properties of the residual signal, core-based bit-plane probabilities are provided for MPEG-4 Audio Scalable to Lossless Coding (SLS), which matches the quantization and coding performed in the core layer. Using the same strategy, new probabilities are obtained to consider the clipping effect in bit-plane coding of an unbounded signal, which is useful for non-core mode of SLS coding. Simulations show that considering the core layer parameters and the clipping effect improve the bit-plane probabilities estimation compared to the existing method.

A scalable audio coding method is proposed using a technique, Quantization Index Modulation, borrowed from watermarking. Some of the information of each layer output is embedded (watermarked) in the previous layer. This approach leads to a saving in bitrate while keeping the distortion almost unchanged. This makes the scalable coding system more efficient in terms of Rate-Distortion. The results show that the proposed method outperforms the scalable audio coding based on reconstruction error quantization which is used in practical systems such as MPEG-4 AAC.

A fine grain scalable coding for audio signals is proposed where the entropy coding of the quantizer outputs is made scalable. By constructing a Huffman-like coding tree where internal nodes can be mapped to reconstruction points, we can prune the tree to control the distortion of the quantizer. Our results show the proposed method improves existing similar work and significantly outperforms scalable coding based on reconstruction error quantization as used in practical systems, eg. MPEG-4 audio.

This paper examines the correlation properties of quantization noise. The quantization noise energy is subtractive if the quantizer output levels are optimized for the probability density of the input signal (pdf optimized). This paper gives a new result that shows that a quantizer (uniform or not) which has quantizer break points midway between output levels (a minimum distance quantizer) and is scaled to minimize the mean-square error, also has this property. Examples are shown that show the correlation properties which determine whether the quantization noise energy is subtractive or additive. This paper also considers a postfilter configuration that compensates for the quantization noise. The postfilter frequency domain gains take the correlation properties of the quantization noise into account. An experiment on reducing the effect of quantization noise in speech gives an indication that taking account of the correlation is useful.

In Voice-over-IP, the quality of interactive conversation is important to users. Quality-based playout buffering seeks an optimum balance between delay and loss. However, such a scheme still suffers when packet losses are bursty. Path diversity can alleviate the effect of losses and improve perceived quality by providing redundancy. In this paper, a new scheme is proposed which evaluates the performance of both paths. We consider three different path diversity schemes. The playout scheduling algorithms are designed based on conversational quality including both calling quality and interactivity. The simulation results show the efficacy of our algorithms in correcting for losses (isolated and burst) and improving perceived conversational quality.

Cette thèse étudie en détail les améliorations apportées au codeur de la parole ITU-T G.711.1. Le codeur original G.711 est en fait un quantificateur μ-law. Le prolongement large-bande G.711.1 utilise le façonnage du bruit ainsi qu'une couche d'amélioration de la bande-basse en plus de la bande-haute. Afin d'améliorer le codage de la bande-basse principale, nous étudions l'utilisation de quantification vectorielle et la décision à retardement. Le codeur arboriforme avec décision à retardée est réalisé par l'algorithme(M,L). Le nouveau quantificateur considère l'information passée et par conséquent, il considère également la propagation de l'erreur engendrée par le façonnage du bruit. Il code plusieurs échantillons par μ-law. Le flot binaire final est compatible avec le décodeur du prolongement large-bande G.711.1 et donc naturellement avec le décodeur du G.711 original. Une méthode d'évaluation, ITU-T P.862 (PESQ) est utilisée pour évaluer la performance. Les résultats montrent que la quantification vectorielle et le codeur arboriforme sont perceptuellement plus performants que le codeur original de la bande principale. Nous notons tout de même qu'ils sont numériquement plus complexes à réaliser. Des études supplémentaires sont suggérées.

In this paper, we extend our previous work on exploiting speech temporal properties to improve Bandwidth Extension (BWE) of narrowband speech using Gaussian Mixture Models (GMMs). By quantifying temporal properties through information theoretic measures and using delta features, we have shown that narrowband memory significantly increases certainty about highband parameters. However, as delta features are non-invertible, they can not be directly used to reconstruct highband frequency content. In the work presented herein, we embed temporal properties indirectly into the GMM structure through a memorydependent tree-based approach to extend representation of the narrow band. In particular, sequences of past frames are progressively used to grow the GMM in a tree-like fashion. This growth approach results in reliable estimates for the GMM parameters such that Maximum Likelihood estimation is no longer necessary, thus circumventing the complexity accompanying high-dimensionality GMM training. Index Terms: Bandwidth extension, GMMs, speech memory. 1.

In Voice-over-IP, buffer delay and packet loss are two main factors effecting perceived conversational quality. A quality-based algorithm aims to seek an optimum balancing of delay versus loss. To improve perceived quality further, steps should be taken to mitigate the effect of losses due to network (missing packets) and buffer underflow (late packets) without increasing buffer delays. In this paper, we propose a quality-based playout algorithm with an FEC design based on conversational quality including calling quality and interactivity. The simulation results show our algorithm’s efficiency of correcting for losses (isolated and burst) and improving perceived conversational quality.

In this paper, we continue our previous work on improving Bandwidth Extension (BWE) of narrowband speech. We have shown that including memory into the parametrization frontend (through delta features) results in higher highband certainty irrespective of feature type, with MFCCs exhibiting higher correlation, in general, between both bands, reaching twice that using LSFs. By incorporating memory into the frontend of a conventional LP-based BWE system, we were able to translate the higher correlation due to memory into BWE performance improvement. Using high-resolution inverse DCT, we also achieved high quality speech reconstruction from MFCCs, thus enabling MFCC-based BWE with improved performance compared to conventional static LP-based BWE. We continue this work by incorporating the superior correlation properties of frontend memory into our MFCC-based BWE system. Log-Spectral Distortion as well as the more perceptually-correlated Itakura-based measures show that incorporating memory into our MFCC-based BWE system results in BWE performance superior to that of our dynamic LP-based BWE system.

We present a novel MFCC-based scheme for the Bandwidth Extension (BWE) of narrowband speech. BWE is based on the assumption that narrowband speech (0.3–3.4 kHz) cor-relates closely with the highband signal (3.4–7 kHz), en-abling estimation of the highband frequency content given the narrow band. While BWE schemes have traditionally used LP-based parametrizations, our recent work has shown that MFCC parametrization results in higher correlation between both bands reaching twice that using LSFs. By employing high-resolution IDCT of highband MFCCs obtained from nar-rowband MFCCs by statistical estimation, we achieve high-quality highband power spectra from which the time-domain speech signal can be reconstructed. Implementing this scheme for BWE translates the higher correlation advantage of MFCCs into BWE performance superior to that obtained using LSFs, as shown by improvements in log-spectral distortion as well as Itakura-based measures (the latter improving by up to 13%). Index Terms: Bandwidth extension, high-resolution IDCT, highband certainty, mutual information, source-filter model

For the purpose of improving Bandwidth Extension (BWE) of narrowband speech, we continue our recent work on the positive effect of exploiting the temporal correlation of speech on the dependence between speech frequency bands. We have shown that such memory inclusion into MFCC speech parametrization translates into higher highband certainty. In the work presented herein, we employ VQ to estimate highband discrete entropies, thus refining our analysis of the effect of memory inclusion on increasing highband certainty. Moreover, we extend our pre-vious analysis to LSF parameters. We further construct a BWE system that exploits our memory inclusion technique, thus trans-lating highband certainty gains into practical BWE performance improvement as measured by the objective quality of recon-structed speech. Results show that memory inclusion decreases the log-Spectral Distortion of the extended highband speech by as much as 1 dB corresponding to more than 14 % relative.

In this paper, we present a speech and audio analysis-synthesis method based on a Basilar Membrane (BM) model. The audio signal is represented in this method by the Hilbert envelopes of the responses to complex gammatone filters uniformally spaced on a critical band scale. We show that for speech and audio signals, a perceptually equivalent signal can be reconstructed from the envelopes alone by an iterative procedure that estimates the associated carrier for the envelopes. The rate requirement of the envelope information is reduced by low-pass filtering and sampling, and it is shown that it is possible to recover a signal without audible distortion from the sampled envelopes. This may lead to improved perceptual coding methods. Index Terms: speech and audio coding, auditory representation, auditory model inversion

In this paper, a novel loading algorithm consisting of four variants is proposed for a multicarrier transceiver employing multiple antennas configured for spatial diversity. The primary objective of the proposed algorithm is to increase the overall throughput while ensuring the mean bit error rate (BER) is below a specified limit. To achieve this, spatial diversity is employed to improve the subcarrier signal-to-noise ratio (SNR) values. Simultaneously, (uniform or non-uniform) bit allocation, which is a function of subcarrier SNR, is performed to increase throughput. To reduce power consumption, spatial processing complexity, and hardware costs, antenna subset selection is also performed by the proposed algorithm to choose a set of active transmit/receive antennas. The results show that combining bit allocation with spatial diversity (employing antenna subset selection) can yield substantial throughput increases.

In this paper, we investigate the effect of temporal correlation on the dependence between the speech narrow and high frequency bands covering the 0.3-3.4 kHz and 3.7-8 kHz ranges, respectively. We follow the technique of using Gaussian mixture modelling of spectral envelopes represented by Mel-frequency cepstral coefficients. The correlation between the disjoint speech frequency bands is quantified through mutual information (MI) and its ratio to highband entropy. Speech exhibits considerable temporal correlation that is not explicitly accounted for by static parametrization of spectral envelopes. Including memory in speech parametrization (through delta features) incorporates such temporal information of speech in its modelling, and hence, MI gains are to be expected resulting in bandwidth extension with better performance. Results show that exploiting delta features can increase certainty about the highband (ratio of MI to highband entropy) by as much as 216% relatively, corresponding to an absolute increase of 12%

Comfort noise insertion during speech pause has been applied to Voice-over-IP and wireless networks for increasing bandwidth efficiency. We present two classified comfort noise generation (CCNG) schemes using Gaussian Mixture classifiers (GMM-C). Our first scheme employs a classified prototype background noise codebook with the prototype noise waveform chosen using a GMM-C. The second scheme utilizes a classified enhanced excitation codebook. The new CCNG algorithms provide better comfort noise during speech pauses and a smaller misclassification rate. We have retrofitted the scheme into existing speech transmission system, such as ITU-T G.711/Appendix II and G.723.1/Annex A. The perceived quality of a voice conversation of the novel system has been noticeably enhanced for car and babble noise. For the G.711 system, a large improvement is obtained for car noise while the largest amelioration is for babble noise in the G.723.1 system.

In last few years, voice over Internet protocol (VoIP) has been gaining popularity as an alternative to traditional telephone by transmitting voice signals as packets over the Internet and private IP-based networks. However, voice packets experience loss, delay, and delay variation, which requires buffering, playout scheduling and loss concealment at the receiver. In this paper, we give an overview of a VoIP application and show how playout scheduling and loss concealment are jointly used to optimize perceived speech quality. We use this optimization criterion to design a histogram-based playout scheduling algorithm. Then, we identify the limitations of the VoIP application for this scheme and propose improvement using path diversity approach that can be implemented via a service overlay network (SON). We present simulations results that show significant improvement of VoIP quality by using this approach

This paper describes a coding paradigm using coding tools based on the characteristics of the human hearing system so as to accommodate a wide range of narrow-band audio inputs without annoying artifacts at low rates (down to 8 kb/s). The narrow-band perceptual audio coder (NPAC) employs a variety of algorithms to account for the perceptually irrelevant parts of the input signal in addition to statistical redundancies. The new algorithms used in the NPAC coder include a perceptual error measure in training the codebooks and selecting the best codewords which takes into account the audible parts of the quantization noise, a perception-based bit-allocation algorithm and a new predictive scheme to vector quantize the scale factors. The NPAC coder delivers acceptable quality without annoying artifacts for most narrow-band audio signals at around 1 bit/sample. Informal subjective tests have shown that the NPAC coder outperforms a commercial low-rate music coder operating at 8 kb/s.

A stream coding framework is presented for solving the distortion-constrained time-frequency dependent quantization problem that naturally arises when overlapped time-frequency decompositions are used. The main contributions of this paper are: (1) an efficient rate-distortion allocation algorithm for dependent quantization when the neighborhood of dependency is large; and (2) demonstration that a perceptual excitation distortion measure produces better coded audio quality than the conventional noise-to-mask ratio measure.

A novel perceptual postfilter is introduced. For each frame, the filter gains, z, are estimated given a vector, y, of the quantized LSFs and the long-term prediction gain of the corresponding frame. The proposed perceptual postfilter is derived from an optimal MMSE estimator, i.e. the estimated gain vector is ˆz = E{z|y}. The MMSE estimator is based on the conditional pdf of z given y, which is computed from the joint pdf modelled by a GMM. The proposed perceptual postfilter improves the speech naturalness comparing with the conventional adaptive postfilter, while maintaining the property of being an “add-on ” postfilter without modification to the current encoder. 1

A voice quality prediction method based on Gaussian mixture models (GMMs) is improved by constructing a feature selection algorithm to provide the best GMM-based prediction quality. The proposed sequential se-lection algorithm performs N-survivor search, allowing for trading between design complexity and performance. Simulation shows that predictors designed using the pro-posed algorithm outperform two benchmark selection al-gorithms. Performance improvements over the ITU-T P.862 PESQ standard are also attained. 1.

We present discrete adaptive bit loading algorithms for multicarrier systems with uniform (nonadaptive) power allocation operating in a frequency selective fading environment. The algorithms try to maximize the overall throughput of the system while guaranteeing that the mean bit error rate (BER) remains below a prescribed threshold. We also study the impact of imperfect subcarrier signal-to-noise ratio information on throughput performance. Results show that the proposed algorithms have approximately the same throughput and mean BER as the optimal allocation while having a significantly lower computational complexity relative to other algorithms with near-optimal allocations. Moreover, when compared with algorithms that employ approximations to water filling, the computational complexity is comparable while the overall throughput is closer to the optimum.

The paper presents a computationally efficient distance measurement method for double binary turbo codes, such as these used in the digital video broadcast with return channel via satellite (DVB-RCS) standard, based on Garello's method. Distance spectra for all standardized DVB-RCS packet sizes and all standardized code rates are presented. A new interleaver design for DVB-RCS based on the dithered relative prime (DRP) interleaving approach is also presented. A minimum distance (d/sub min/) of 36 has been achieved for an unpunctured ATM packet of 424 information bits with a DRP interleaver, whereas the d/sub min/ of the standardized DVB-RCS interleaver is 31. A d/sub min/ of 38 has been achieved for an unpunctured MPEG packet of size 1504 information bits with a DRP interleaver, whereas the d/sub min/ of the standardized DVB-RCS interleaver is 33. Simulation results for code rate 1/3 show an improvement at high signal to noise ratios of at least 0.15 dB and 0.25 dB for ATM and MPEG packets, respectively.

In this paper, we evaluate and compare the robustness of several adaptive bit loading algorithms for multicarrier transmission systems, when imperfect subcarrier signal-to-noise ratio (SNR) information is used. In particular, we investigate the impact of the uncertainty of data-aided channel estimation techniques on system performance. We also examine an implementation issue associated with adaptive bit loading algorithms that use metrics related to the SNR. Although such metrics can be derived via closed form expressions, look-up tables are used instead to reduce system complexity, resulting in the SNR values being quantized. Thus, we examine the effects of SNR quantization on system performance. Finally, we present a technique for choosing SNR values in a fixed length look-up table in order to minimize quantization error.

We examine the problem of bit allocation when time spread and frequency spread perceptual distortion criteria are used. For such measures, standard incremental techniques can fail. Two algorithms are introduced for bit allocation; the first a multi-band version of the greedy algorithm, and the second an inverse greedy algorithm initialized by the bit allocation of a forward algorithm driven by a non-spread metric. Experimental results show the second algorithm outperforms the first.

Current public telephone networks compromise voice quality by bandlimiting the speech signal. Telephone speech is characterized by a bandpass response from 300 to 3400 Hz. The voice quality is perceived as being much worse than for wideband speech (50-7000 Hz). We present a novel approach which combines equalization and estimation to create a wideband signal, with reconstructed components in the 3400 Hz to 7000 Hz range. Equalization is used in the 3400-4000 Hz range. Its performance is better than statistical estimation procedures, because the mutual dependencies between the narrowband and highband parameters are not sufficiently large. Subjective evaluation using an improvement category rating shows that the reconstructed wideband speech using both equalization and estimation substantially enhances the quality of telephone speech. We have also evaluated the performance on the narrowband output of several standard codecs. Overall, the use of equalization for part of the highband regeneration makes the system more robust to phonetic variability and speaker gender.

Abstract—An algorithm for nonintrusive speech quality esti-mation based on Gaussian mixture models (GMMs) is presented. GMMs are used to form an artificial reference model of the behavior of features of undegraded speech. Consistency measures between the degraded speech signal and the reference model serve as indicators of speech quality. Consistency values are mapped to an objective speech quality score using a multivariate adaptive regression splines function. When tested on unseen data, the proposed algorithm generally outperforms ITU-T standard P.563, which is the current “state-of-the-art ” algorithm. The algorithm computes objective quality scores roughly twice as fast as P.563. Index Terms—Gaussian mixtures, quality assurance, quality measurement, quality of service, speech coding, speech quality, speech transmission, telephony. I.

Traditional teleconferencing uses a select-and-mix function at a centralized conferencing bridge.In VoIP environments, this mixing operation can lead to speech degradation when using high compression speech codecs due to tandem encodings and coding of multi-talker signals.A tandem-free architecture can eliminate tandem encodings and preserve speech quality.VoIP conference bridges must also consider the variable network delays experienced by different packetized voice streams.A synchronized speaker selection algorithm at the bridge can smooth out network delay variations and synchronize incoming voice streams.This provides a clean mapping of the N input packet streams to the M output streams representing selected speakers.This paper presents a synchronized speaker selection algorithm and evaluates its performance using a conference simulator.The synchronization process is shown to account for only a small part of the overall delay experienced by selected packets.

The bandwidth for telephony is generally defined to be from 300–3400 Hz. This bandwidth restriction has a no-ticeable effect on speech quality. We present an algo-rithm which recovers the missing highband parts from telephone speech. We describe an MMSE estimator us-ing hard/soft-classification to create the missing high-band spectrum envelope. The classification is motivated by acoustic phonetics: voiced vowels and consonants, and unvoiced phonemes demonstrate different character-istic spectra. The classification also captures gender dif-ferences. A hard classification on phoneme character-istic parameters, such as a voicing degree and a pitch lag, reduces the MMSE of the highband spectrum en-velope estimates. An estimator using HMM-based soft-classification can further bring down the estimated high-band spectrum distortion by taking the time evolution of the spectra into consideration. Objective measures (mean log-spectrum distortion) and spectrograms con-firm the improvement noted in informal subjective tests. 1

We present an efficient bit allocation algorithm for multicarrier systems operating in frequency-selective environment. The proposed algorithm strives to maximize the overall throughput while guaranteeing that the mean bit error rate (BER) remains below a prescribed threshold. The algorithm is compared with several other algorithms found in literature in terms of the overall throughput, mean BER, and relative computational complexity. Furthermore, the algorithms are compared with an exhaustive search routine to determine the optimal bit allocation in terms of maximizing throughput given the constraint on error performance. No power allocation is performed by the algorithms. Results show that the proposed algorithm has approximately the same throughput and mean BER as the optimal solution while possessing a significantly lower computational complexity relative to the other algorithms with similar performance. When compared to algorithms which employ approximations to waterfilling, the computational complexity is comparable while the overall throughput is closer to the optimum.

A new frequency domain approach to coding audio signals is introduced. The bit assignment strategy is aimed at reducing the perceived loudness difference between the original signal and the coded signal. As such it uses perceptual effects (spread excitation patterns), but does not directly invoke masking results. At low bit rates, examples coded with the new approach sound better than a more traditional bit allocation based on noise-to-mask ratio.

This paper examines schemes that modify linear prediction (LP) analysis for speech signals. First, techniques which improve the conditioning of the LP equations are examined. White noise compensation for the correlations is justified from the point of view of reducing the range of values which the predictor coefficients take on. The efficacy of the procedure is measured over a large speech database. Various techniques for bandwidth expansion of the LP spectral peaks are also examined. These include lag windowing of the correlation, windowing of the predictor coefficients, and modification of the line spectral frequencies. New formulas for the bandwidth expansion factor are given.

Traditional telephone conferencing has been accomplished by way of a centralized conference bridge. The tandem arrangement of high compression speech codecs in conventional VoIP conference bridges lead to speech distortions and require a substantial number of computations. Decentralized architectures avoid the speech degradations and delay, but lack strong control and depend on silence suppression to make the endpoint bandwidth and processing requirements scalable. One solution is to use centralized speaker selection and forwarding, and decentralized decoding and mixing. This approach eliminates the problem of tandem encodings but maintains centralized control, thereby improving the speech quality and scalability of the conference. This paper considers design options and solutions for this model in the context of modern IP telephony networks. Performance was evaluated with real conferees over live conferences using a PC-based bridge and a third-party endpoint. Conferees strongly preferred the speech quality of the new arrangement to that of a conventional VoIP conference bridge.

ABSTRACT: Speech sounds occupy 8 kHz or more of bandwidth. However, current public telephone networks limit the speech bandwidth to 300–3400 Hz. Telephone speech is characterized by thin and muffled sounds, and degraded speaker identification. We describe an algorithm which generates the missing highband components from the narrowband speech signal. The algorithm is based on three acoustic-phonetic classified narrowband-to-wideband linear prediction (LP) spectrum mapping codebooks to recover the missing highband spectrum. Subjective tests show that the reconstructed wideband speech improves the speech quality. LP spectrum bandwidth expansion is used to avoid sharp spectral peaks. The mean SD (log-spectrum distortion) decreases by 0.93 dB, comparing to non-classified codebooks without LP bandwidth expansion.

A methodology for suppressing musical noise produced by signal subspace speech enhancement is presented. An auditory post-filter is placed at the output of the subspace filter to smooth the enhanced speech spectra. By utilizing a perceptual filter, averaging is performed in a manner similar to that of the human auditory system. As such, distortion to the underlying speech signal is reduced.

Conventional interpolation of Linear Prediction (LP) parameters can provide a poor spectral match to the underlying speech signal for the intermediate subframes. In this paper, we present a method of modifying the interpolation endpoint LP parameters to improve the spectral tracking over all subframes. This ‘warping’ algorithm is based on minimizing the distortion between the interpolated LP parameters and those computed directly from the speech signal. The algorithm has been integrated into the Adaptive Multi Rate (AMR) speech codec. Our results show that this method enhances coder performance by smoothing out the LP parameter tracks and reducing coding distortion.

Conventional linear prediction (LP) techniques can fail to adequately model speech spectra when the model order is too low and/or when the input is periodic (voiced speech). We view the LP modelling problem as a correlation matching problem. We introduce a correlation matching criterion which models the signal as a filtered mixture of a noise-like excitation and a periodic excitation. As such it is an extension of the discrete all-pole (DAP) modelling approach. The new technique provides a means to generate LP spectra that evolve more smoothly from frame to frame even when the excitation signal has a periodic component with changing period.

This paper presents a bi-modal coding paradigm to com-press narrowband audio signals at 8 kbit/s. In the general mode, the Enhanced Narrowband Audio Coder (ENPAC) exploits the characteristics of the human hearing system to adaptively code the perceptually important spectral components of the input audio. The other mode is em-ployed to handle audio inputs with a strong harmonic structure. In that mode, the input block is represented by its audible harmonics. The spectral magnitude is mod-eled by the linear prediction analysis in the time domain. The phase of each harmonic is predicted and the phase residues are quantized using an adaptive bit allocation algorithm. This paper introduces a perceptually-based upper bound for phase errors of spectral components. The ENPAC encoder delivers good quality for narrow-band speech and non-speech inputs. 1

A non-maximally decimated filterbank structure for pre-equalizing channels with intersymbol interference (ISI) is investigated. The impulse response of the channel is assumed to be known at the transmitter. Compared with the classical Tomlinson-Harashima (1971) precoding technique, the proposed pre-equalizer compensates for the channel without increasing the number of the received signal levels (channel alphabet). The proposed technique does not require the channel to be minimum-phase. The filterbank structure adds redundancy to the input signal to compensate for the channel ISI while keeping the transmitted power bounded. The proposed pre-equalization is particularly useful for data transmission over voice-band PCM channels. The upstream PCM channel is bandlimited, causing severe ISI at the output of the front-end receiver filter. By using the pre-equalizer at the transmitter, channel ISI can be mitigated.

Automatic discrimination of speech and music is an important tool in many multimedia applications. Previous work has focused on using long-term features such as differential parameters, variances and time-averages of spectral parameters. These classifiers use features estimated over windows of 0.5-5 seconds, and are relatively complex. We present our results of combining the line spectral frequencies (LSFs) and zero crossing-based features for frame-level narrowband speech/music discrimination. Our classification results for different types of music and speech show the good discriminating power of these features. Our classification algorithms operate using only a frame delay of 20 ms, making them suitable for real-time multimedia applications.

In this work we consider adaptive bit allocation for perceptual coding of narrowband audio signals at low rates (down to 8 kbit/s). Two different strategies are used to shape the audible noise spectrum. In one approach, the quantization noise spectrum is shaped in parallel with the masking threshold curve. This way the noise is equally audible in different frequency bands. The other approach generates a flat noise spectrum above the masking threshold. The noise power is not equally distributed over the frequency range, hence it is audible to various extents at different frequencies.

Background environmental noises degrade the performance of speech-processing systems (e.g. speech coding, speech recognition). By modifying the processing according to the type of background noise, the performance can be enhanced. This requires noise classification. In this paper, four pattern-recognition frameworks have been used to design noise classification algorithms. Classification is done on a frame-by-frame basis (e.g. once every 20 ms). Five commonly encountered noises in mobile telephony (i.e. car, street, babble, factory, and bus) have been considered in our study. Our experimental results show that the line spectral frequencies (LSFs) are robust features in distinguishing the different classes of noises.

This paper discusses perceptual coding of narrowband audio signals at low rates. In particular, it proposes a new error measure which shapes the noise inside the critical bands, a window switching criterion based on the temporal masking effect of the hearing system, a more accurate model of the simultaneous masking effect of the hearing system, perceptually-based bit allocation algorithms based on two different approaches towards quantization noise shaping and a predictive vector quantization scheme to code the scale factors. The resulting coding scheme outperforms existing low rate speech coders for non-speech signals.

Acoustic echo cancellation can be used to remove the annoying talker feedback in hands-free (teleconferencing) systems. In this thesis, we examine the DCT-LMS algorithm and we emphasize on the role played by the Discrete Cosine Transform. This fixed transformation reduces the eigenvalue spread of the input autocorrelation matrix by partially decorrelating the inputs. The autocorrelation matrix of speech signals is often rank-deficient. During the low rank phases, some of the transform-domain tap coefficients become irrelevant to the adaptation process and stop adapting. When the autocorrelation matrix gains full rank, there will be no longer any "frozen" weights. However, the weights that have been frozen are "far" from the optimal point; they require additional convergence time to track again the changes in the room impulse response. In this dissertation, we present a new method that uses the information contained in the other coefficients to move the frozen weights closer to the optimal point and, consequently, reduce the overall convergence time. (Abstract shortened by UMI.)

Existing approaches to background noise coding at very low bit rates (i.e., below 1 kbps) fail to reproduce the noise with natural quality, resulting in a degradation of the overall perceived quality. In this paper, we propose a novel scheme for natural-quality reduced-rate coding of background acoustic noise in voice communication sys-tems. A better representation of the excitation signal in a noise-synthesis model is achieved by classifying the type of acoustic environment noise. Class-dependent residual substitution is used at the receive side to synthesize back-ground noise that sounds similar to the background noise at the transmit side. The improvement in the quality of synthesized noise during speech gaps helps in preserving noise continuity between talk spurts and speech pauses, and enhances the perceived quality of a conversation. 1

Data transmission over bandlimited channels requires pulse shaping to eliminate or control intersymbol interference (ISI). Nyquist filters provide ISI-free transmission. Here we introduce a phase compensation technique to design Nyquist filters. Phase compensation can be applied to the square-root of any zero-phase bandlimited Nyquist filter with normalized excess bandwidth less than or equal to one. The resulting phase compensated square-root filter is also a Nyquist filter. In the case of a raised-cosine spectrum, the phase compensator has a simple piecewise-linear form. Such a technique is particularly useful to accommodate two different structures for the receiver, one with a filter matched to the transmitting filter and one without a matched filter. We also use the phase compensation technique to characterize a more general family of Nyquist filters which subsumes raised-cosine spectra. These generalized raised-cosine filters offer more flexibility in filter design. For instance, the rate of asymptotic decay of the filter impulse response may be increased, or the residual ISI, introduced by truncation of the impulse response, may be minimized. Design examples are provided to illustrate these choices.

This article discusses the problem of optimizing a lattice-based signal constellation for signaling over a partial response channel. The objective is to minimize the probability of error which is determined by the combined effects of the additive Gaussian noise and channel memory.

Linear prediction (LP) coefficients are used to describe the formant structure of a speech waveform. Many factors contribute to the frame-to-frame fluctuation of these parameters. These variations adversely affect the performance of the LP quantizer and the quality of the synthesized speech. For voiced speech, efficient coding of the pitch pulses at the output of the inverse formant filter relies on the similarity of successive pitch waveforms. The performance of this coding stage is also jeopardized by LP variations. In this paper, we propose a new method which smooths the evolution of the LP parameters. Our algorithm is based on matching the output of the formant predictor to a target signal constructed using smoothed pitch pulses. With this approach we have successfully reduced the frame-to-frame variation of LP coefficients, while increasing the similarity of pitch pulses.

Nonlinear predictive split vector quantization (NPSVQ) and classified NPSVQ (CNPSVQ) are introduced to exploit the correlation among the speech spectral parameters from two adjacent analysis frames. By interleaving intraframe SVQ with forward predictive SVQ, the error propagation is limited to at most one adjacent frame. At an overall bit rate of about 21 bits/frame, NPSVQ can provide similar coding quality as intraframe SVQ at 24 bits/frame. Voicing classification as used in CNPSVQ to obtain an additional average gain of 1 bit/frame for unvoiced frames. Therefore, an overall bit rate of 20 bits/frame is obtained for unvoiced frames. The particular form of nonlinear prediction we use incurs virtually no additional encoding computational complexity. We have verified our comparative performance results using subjective listening tests.

We introduce an efficient addressing scheme for the nearly optimum shaping of a multidimensional signal constellation. The 2-D (two-dimensional) subspaces are partitioned into K energy shells of equal cardinality. The average energy of a 2-D shell can be closely approximated by a linear function of its index. In an N=2n-D space, we obtain K/sup n/ shaping clusters of equal cardinality. Shaping is achieved by selecting T/spl les/K/sup n/ of the N-D clusters with the least sum of the 2-D indices. This results in a set of T integer n-tuples with the components in the range [0, K-1] and the sum of the components being at most a given number L. The problem of addressing is to find a one-to-one mapping between the set of such n-tuples and the set of integers [0, T-1] such that the mapping and its inverse can be easily implemented. In the proposed scheme, the N-D clusters are grouped into blocks of identical binary weight vectors. This results in a simple rule for the addressing of points within the blocks. The addressing of the blocks is based on some recursive relationship which allows us to decompose the problem into simpler parts. The overall scheme requires a modest amount of memory and has a small computational complexity.

A method is developed to compute the weight distribution of a set of points obtained from a lattice. The lattice is scaled (with possibly nonequal factors) along different dimensions, is shifted to an arbitrary point, and its lower dimensional subspaces are truncated within given shaping regions. Each branch in the lattice trellis diagram is labeled by the weight distribution of the corresponding coset incorporating the effects of scaling, shifting, and truncation. The weight distribution is obtained by multiplying the weight distribution of the serial branches and then adding the result over parallel paths.>

We find analytical expressions for the block-based input and output eigenvectors and eigenvalues of the systems with responses 1/spl plusmn/d and 1-D/sup 2/. The input eigenvectors form an orthonormal basis which is the optimum modulator for a channel with that transfer function. The output eigenvectors form an orthonormal basis with the same spectral nulls as the corresponding system. This basis can be used to produce line codes with spectral nulls. The eigenvectors are sinusoids. This reduces the computational complexity by allowing for fast transform algorithms to perform the modulation for a block of data.>

This paper describes a lookup table for the addressing of an optimally shaped constellation. The method is based on partitioning the subconstellations into shaping macro-shells of integer bit rate and increasing average energy. The macro-shells do not need to have an equal number of points. A lookup table is used to select a subset of the partitions in the cartesian product space. By devising appropriate partitioning/merging rules, we obtain suboptimum schemes of very low addressing complexity and small performance degradation. The performance is computed using the weight distribution of an optimally shaped constellation.>

The structure of the regions which provide the optimum tradeoff between gamma /sub s/ (shaping gain) and CER (constellation-expansion ratio) and between gamma /sub s/ and PAR (peak to average power ratio) in a finite dimensional space is introduced. Analytical expressions are derived for the corresponding tradeoff curves. In general, the initial parts of the curves have a steep slope. This means that an appreciable portion of the maximum shaping gain, corresponding to a spherical region, can be achieved with a small value of CER/sub s/, PAR. The technique of shell mapping is introduced. This is a change of variable which maps the optimum shaping region to a hypercube truncated within a simplex. This mapping is a useful tool in computing the performance, and also in facilitating the addressing of the optimum shaping region. Using the shell mapping, a practical addressing scheme is presented that achieves a point on the optimum tradeoff curves. For dimensionalities around 12, the point achieved is located near the knee of the corresponding tradeoff curve. For larger dimensionalities, a general shaping region with two degrees of freedom is used. This region provides more flexibility in selecting the tradeoff point.>

For Pt. I, see ibid., pp. 1799-1808, Nov. 1993. By appropriately selecting the boundary of a multidimensional signal constellation used for data transmission, the average energy of the constellation can be reduced. Reduction in the average energy (shaping gain) is obtained at the price of increasing the constellation-expansion ratio (CER/sub s/) and the peak-to-average-power ratio (PAR). The authors describe some practical means for selecting the boundary so as to achieve a point with low addressing complexity near the knee of the corresponding tradeoff curves (shaping gain versus CER/sub s/ or PAR). One class of addressing schemes is based on using a lookup table. A method to facilitate the realization of the addressing lookup table is introduced. This method is based on the decomposition of addressing into a hierarchy of addressing steps, each of a low complexity. This avoids exponential growth of the complexity. Using this addressing decomposition and a memory of a practical size, one can move along a tradeoff curve which has negligible suboptimality. Another class of addressing schemes is based on using a Voronoi constellation in a space of half the original dimensionality.>

The authors try to identify the primary sources of distortion in a non-recursive time-scale modification (TSM) algorithm which is based on the short-time Fourier transform (STFT). A simpler version of this TSM algorithm is then proposed for processing speech, where incremental estimators eliminate the need for explicit linear time-scaling operations. Also featured in the design is a waveform structure compensation stage to prevent excessive deterioration of the rate-changed output. A polar (i.e., magnitude-phase) synthesis equation is used for increased efficiency. The TSM method is capable of generating high-quality rate-changed speech at a reasonable computational cost.>

A new algorithm for the processing of signals by an array of sensors is proposed. The objective is to find the number and the directions of arrival (DOA) of signals impinging on a linear array. The predictive stochastic complexity criterion of Rissanen (1986) is used to select the best model order. To reduce the computational load, the algorithm operates with a suboptimal estimator while maintaining the consistency of the estimator. The proposed method is on-line and can be utilized in time-varying systems for target tracking. The method can be used for both correlated and uncorrelated signals.>

We view the speech enhancement task in two aspects: reduction of the perceptual noise level in degraded speech and reconstruction of the degraded information, which may result in improvement of speech intelligibility. We are also very interested in noiseindependent speech enhancement where test noise environments could differ in intensity from those of algorithm development. To this end, we have developed in this paper an algorithm called Noise-Independent Statistical Spectral Mapping (NISSM) to estimate a speech enhancement Wiener filter. NISSM consists of a noise-resistant transformation, which converts noisy speech to a set of noise-resistant features, and a spectral mapping function, which maps the features to autoregressive spectra of clean speech. We will show that the proposed algorithm effectively reduces noise intensity. When the noise intensity of training differs from that of testing, NISSM outperforms significantly a conventional spectral mapping. The algorithm operates frame-by-frame and is designed for real-time application. The noise interference could be stationary or non-stationary white noise with variable intensity. I.

Complex cepstral deconvolution is applied to acoustic dereverberation. It is found that traditional cepstral techniques fail in acoustic dereverberation because segmentation errors in the time domain prevent accurate cepstral computation. An algorithm for speech dereverberation which incorporates a novel approach to the segmentation and windowing procedure for speech is presented. Averaging in the cepstrum is exploited to increase the separation between the speech and impulse response. An estimate of the room impulse response is built, and a least squared error inverse filter is used to remove the estimated impulse response from the reverberant speech. Reduction of reverberation with this technique is demonstrated.>

A stochastic tree coder based on the (M,L) search algorithm suggested by V. Iyengar and P. Kabal (1988) and a low-day CELP (code-excited linear prediction) coder proposed by J.H. Chen (1989) are considered. The individual components (predictors, gain adaptation, excitation coding) of the two coders are analyzed. The performances of the two types of coders are compared. The two coders have comparable performance at 16 kb/s under clean channel conditions. Methods to improve the performance of the coders, particularly with a view to bringing the bit rate to below 16 kb/s, are studied. Suggestions for improving the performance include an improved high-order predictor (applicable to both coders), and training of the excitation dictionary as well as a better gain adaptation strategy for the tree coder.>

An analysis and simulation results are presented comparing the performance of several types of high-order backward adaptive predictors with orders up to 100. Issues in high-order linear predictive coding (LPC) analysis, such as analysis methods, windowing, ill-conditioning, quantization noise effects, and computational complexity, are studied. The performance of the various analysis methods is compared with the conventional sequential formant-pitch predictor. The auto-correlation method (50th order) shows performance advantages over the sequential formant-pitch configurations. Several new backward high-order methods using covariance analysis and a lattice formulation show much better prediction gains than the auto-correlation method.>

Methods of improving vector-scalar quantization of linear predictive coding (LPC) coefficients with 20 to 30 bits per 20 ms speech frame are studied. The first approach used is to couple the vector and scalar quantization stages. The second innovation is the incorporation of a small adaptive codebook to the larger fixed codebook. Frame-to-frame correlation of the LPC coefficients is exploited at no extra cost in bits. It is shown that the performance of vector-scalar quantization of speech with the use of the two new techniques is better than that of scalar coding techniques currently used in LPC coders.>

The authors investigate the use of CELP (code excited linear prediction) in coding wideband speech signals at an operating rate of 16 kb/s. The wideband signals under consideration are bandlimited to 7500 Hz and sampled at 16 kHz. In order to achieve a low operating rate, the coding places more emphasis on the lower frequencies (0 to 4 kHz), while the higher frequencies are coded less precisely, but with little perceived degradation. To this effect, the basic CELP model is modified to operate in a split-band mode. For the operating rate of 16 kb/s, subjective tests showed that the wideband speech coder was preferred to a high-quality narrowband coder. This wideband speech coder offers an attractive alternative to conventional narrowband codes at rates near 16 kb/s for many applications.>

A set of conditions which allow bandwidth efficient transmultiplexers to be synthesized is developed. The synthesis procedure is based on a generalized impulse response for the combining (modulating) and separation (demodulating) filters. In particular, the combining and separation filters are bandpass versions of one or two lowpass prototypes and are configured to cancel crosstalk by exploiting relationships between the center frequencies, delays, and phases in their impulse response. Based on the derived conditions, five different transmultiplexers are synthesized. Three of them implement orthogonal quadrature amplitude modulation (OQAM) with repeated center frequencies. The other two accomplish vestigial sideband modulation (VSB) with distinct frequencies. The transmultiplexers can be converted into new complementary subband systems. Intersymbol interference is eliminated in both the transmultiplexers and the subband systems by appropriately designing the prototypes.>

The generation of zero-phase FIR (finite impulse response) low-pass Nyquist filters that can be split into minimum- and maximum-phase parts is considered. Two factorable minimax design methods are given. These methods use the McClellan-Parks algorithm as a first step to control the stopband behavior. The time domain constraints, imposed by solving a linear system of equations, determine the passband response. The final filter exhibits equiripple stopband behavior. The advantages of these methods are that the minimum- and maximum-phase parts are obtained without direct factorization and that arbitrary frequency weighting can be easily incorporated to allow for a nonequiripple behavior. The design approach is compared with other methods, and a practical design that conforms to a CCITT voiceband modem specification is shown.>

The authors study a speech coder using a tree code generated by a stochastic innovations tree and backward adaptive synthesis filters. The synthesis configuration uses a cascade of two all-pole filters-a pitch (long time delay) filter followed by a formant (short time delay) filter. Both filters are updated using backward adaptation. The formant predictor is updated using an adaptive lattice algorithm. The multipath (M,L) search algorithm is used to encode the speech. A frequency weighted error measure is used to reduce the perceptual loudness of the quantization noise. The speech coder has low delay (1 ms) and has been evaluated through formal subjective testing to have speech quality that is equivalent to that for 7-bit log-PCM.>

Code-excited linear prediction (CELP) is examined for the coding of speech signals. The methodology is developed to allow for a joint optimization of waveform selection, waveform scaling, and pitch filter determination. Methods to accommodate high-pitch speakers (pitch lag smaller than the analysis frame size) are given. The requirements for coding the synthesis parameters into a bit stream at 4.8 kb/s are discussed. The coder has been tested at channel error rates of 0.001, with only minor degradations in the resulting speech. An adaptive postfilter has been added to achieve a small increase in perceived speech quality.>

Jointly optimized solutions for formant and pitch predictors are discussed. The filters are implemented in transversal form with the formant filter always preceding the pitch (F-P cascade). The jointly optimized methods are further subdivided into a combined solution and an iterative sequential solution. Both yield higher prediction gains than their conventional sequentially determined counterpart. A practical version of the iterated sequential approach, in which the filters are constrained to be minimum phase, generates decoded speech of high perceptual quality when applied in a coding environment.>

The calculation of the discrete Fourier transform using a fast Fourier transform (FFT) algorithm with fixed-point arithmetic is considered. The input data is scaled to prevent overflow and to maintain accuracy. The implementation uses 16-bit fixed-point representation for the data and provides for double precision accumulation of sums and products. Algorithm variants as well as different rounding options are compared. Execution times for implementations based on a single chip signal processor are given. These show that a considerable increase in accuracy can be obtained with only a small penalty in execution time, by applying an alternating form of rounding rather than truncation.

This paper discusses the analysis techniques used to derive the excitation waveform for multipulse coding of speech. A computationally efficient formulation is derived for both covariance and correlation type analyses. These methods differ in the way block edges are treated. Several methods for pulse amplitude and position determination are given, ranging from a purely sequential one to one which reoptimizes pulse amplitudes at each step. It is shown that the reoptimization scheme has a nested structure that allows a reduction in the computations. An efficient method for pulse position coding is given. This method can essentially achieve the entropy limit for randomly placed pulses. Experimental results are given for typical configurations including computational requirements and speech quality assessments.

The necessary and sufficient conditions for the stability and convergence of adaptive minimum mean square error equalizers are developed. The analysis includes the effect of delays in the adjustment path, as occur, for instance, when the decoder in a data transmission system delays its output decisions.

This paper presents a unified study of partial-response signaling (PRS) systems and extends previous work on the comparison of PRS schemes. A PRS system model is introduced which enables the investigation of PRS schemes from the viewpoint of spectral properties such as bandwidth, nulls, and continuity of derivatives. Several desirable properties of PRS systems and their relation to system functions are indicated and a number of useful schemes, some of them not previously analyzed, are presented. These systems are then compared using as figures of merit speed tolerance, minimum eye width, and signal-to-noise ratio (SNR) degradation over ideal binary transmission. A new definition of speed tolerance, which takes into account multilevel outputs and the effect of sampling time, is introduced and used in the calculation of speedtolerance figures. It is shown that eye width, a performance measure that has not been used previously in comparing PRS systems, can be calculated analytically in many cases. Exact values as well as bounds on the SNR degradation for the systems under consideration are presented. The effect of precoding on system performance is also analyzed.