Modern digital signal processing increasingly leverages multi-channel systems and algorithms to tackle emerging challenges, e.g., in beamforming, crosstalk cancellation (CTC) or active noise control. Commonly, the modeling, design and simulation of these systems involves the use of multiple-input multiple-output (MIMO) finite impulse response (FIR) filters. To increase the practical feasibility, these potentially resource-intensive FIR filters can be approximated by infinite impulse response (IIR) filters of lower order. Balanced truncation (BT) can be used for the approximation, but is not used as widely in these areas. A possible reason for this is the state space formulation of the algorithm, which does not scale well for general systems of high order. In this contribution, we present a BT algorithm which is specifically tailored to the approximation of MIMO FIR filters which avoids most of the usually needed computations. This algorithm can be applied to a number of current problems and we study two examples in detail. We find the numerical accuracy to be satisfactory even with high-order systems.

Measuring personal head-related transfer functions (HRTFs) is essential in binaural audio. Personal HRTFs are not only required for binaural rendering and for loudspeaker-based binaural reproduction using crosstalk cancellation, but they also serve as a basis for data-driven HRTF individualization techniques and psychoacoustic experiments. Although many attempts have been made to expedite HRTF measurements, the rotational velocities in today’s measurement systems remain lower than those in natural head movements. To cope with faster rotations, we present a novel continuous HRTF measurement method. This method estimates the HRTFs offline using a Kalman smoother and learns state-space parameters, including the system model, on short signal segments, utilizing the expectation maximization algorithm. We evaluated our method in simulated single-channel and multi-channel measurements using a rigid sphere HRTF model. Comparing with conventional methods, we found that the system distances are improved by up to 30 dB.

This paper presents a deep neural network (DNN)-based system for phase reconstruction of speech signals solely from their magnitude spectrograms. The phase is very sensitive to time shifts. Therefore it is meaningful to estimate the phase derivatives instead of the phase directly, e.g., using DNNs and then apply a phase reconstruction method to recombine these estimates to a suitable phase spectrum. In this paper, we propose three changes for such a two-stage phase reconstruction system. For estimating phase derivatives using DNNs, we propose 1) a new regularized cosine loss function and 2) a preprocessing step for the phase derivatives that eliminates systematic offsets in the data. The experimental results demonstrate that these two changes reduce the training duration and stabilize the training with respect to hyperparameter variations. Furthermore, we propose 3) a new simple but effective averaging of weighted derivative estimates for the subsequent phase reconstruction method. Experimental results confirm that the proposed reconstruction method surpasses state-of-the-art systems both in terms of the reconstructed phase spectra and the overall achieved speech quality.

Crosstalk cancellation filters play a crucial role when reproducing binaural signals over loudspeakers. Least-squares design methods for crosstalk cancellation filters are commonly formulated either in the time domain or in the frequency domain. While the time-domain problem is too complex to be solved in a dynamic real-time setup, the frequency-domain problem is computationally cheap. However, it yields circular convolution artifacts, which degrade the performance. Using a unified framework, we derive a causality-constrained frequency-domain least-squares problem, which itself is an unconstrained problem. Our method does not yield circular convolution artifacts since the impulse responses are causal by construction. This construction is based on the fact that the real part and the imaginary part of a causal signal's spectrum are related via Hilbert transform. By approximating the Hilbert transform, we reduce the proposed method's computational complexity. The complexity reduces by two to three orders of magnitude while the performance remains close to that of the time-domain solution. We provide a detailed complexity analysis and compare our method to three state-of-the-art methods on a large dataset. Our investigations reveal that the causality-constrained method enables a flexible tradeoff between computational complexity and performance figures.

Active noise cancellation (ANC) technology is a valuable asset for hearables. For a well performing and robust ANC system precise knowledge of the relevant acoustic paths is vital. It is feasible to individually measure the user’s secondary path by using the inner loudspeaker and microphone that are already built in into most ANC headphones. For measuring the primary path, however, an external loudspeaker as well as a suitable environment for acoustic measurements is required.We exploit that changes in the primary and secondary path appear jointly, based on the headphone fitting and the physiology of the user’s ear. In this contribution, we propose an estimator for the individual primary path based on a measurement of the individual secondary path. We evaluate the estimator by comparing the ANC performance to conventional fixed filter design methods for a large data set of real acoustic paths of an in-ear ANC headphone.

This paper presents a novel 3DoF+ system that allows to navigate, i.e., change position, in scene-based spatial audio content beyond the sweet spot of a Higher Order Ambisonics recording. It is one of the first such systems based on sound capturing at a single spatial position. The system uses a parametric decomposition of the recorded sound field. For the synthesis, only coarse distance information about the sources is needed as side information but not the exact number of them. In an experiment with multiple sources and reverberation, the proposed method shows good performance in terms of an instrumental measure and excellent performance in a subjective evaluation based on a MUSHRA listening test. A variant of the analysis stage increases localization performance and scores median subject ratings between 91 % and 99 % compared to the reference.

The identification of transversal filters is important for numerous applications. In acoustical applications a first order Markov model is often used to describe the time-variant nature of transversal filters. The Kalman filter is the optimal unbiased estimator for such a Markov model. It inherently calculates the uncertainty of the current state estimate by its state error covariance matrix. In contrast to the broadband Kalman filter the covariance matrix of the exact Kalman filter depends on properties of the input signal. The single step covariance update of the exact Kalman filter is a discrete-time algebraic Riccati equation. We propose a solution for the steady state covariance matrix, which depends on the process parameters of the Markov model as well as properties of the input signal. It is derived based on the eigendecomposition of the covariance matrix and the autocorrelation matrix of the input signal. The proposed algorithm converges in few iterations and gives accurate results. We show how this result can be used to predict the steady state performance of the Kalman filter for system identification through numerical examples.

The headphones market is expanding and new applications based on active noise control (ANC) are emerging. In this contribution, we derive a fixed feed-forward FIR filter for ANC headphones based on a minimum mean-square error (MMSE) cost function that realizes user-defined active equalization and attenuation of the ambient sound. By analyzing the limit of this solution with respect to active attenuation, we derive a generalized solution that allows the user to influence the equalization even for high attenuation. We investigate the importance of a priori information and runtime measurements of the acoustic transfer functions, and show how this can be utilized to further increase the accuracy of the equalization.

A novel approach to non-uniform spatial resolution encoding and dimensionality reduction of Higher Order Ambisonics spatial audio content is proposed. Space warping is applied to modify the spatial arrangement of the audio scene such that the coefficient energy is packed in few low order coefficients. The underlying idea is to squeeze or stretch parts of the spherical coordinates. In contrast to conventional Mixed Order Ambisonics, our approach is based on a transformation. Unlike methods for dimensionality reduction such as PCA, the resulting signal remains a valid HOA signal and spatial relations are maintained. We present a mathematical framework for space warping, demonstrate the potential of the method using an example, and discuss prerequisites and applications.

Mobile devices for hands-free speech communication often show significant nonlinear distortion in the sound emitted by their loudspeakers. Therefore, conventional linear echo cancellation is not sufficient for maintaining a high conversation quality. In this work we propose a nonlinear echo canceller that uses two serially cascaded adaptive filters to compensate for the nonlinear and linear echo. We show that a stable operation of the cascaded structure is achieved by using the multi-channel Kalman algorithm in the frequency domain with filtered-x references. By modelling a nonlinearity with memory we further improve the performance. To reduce the computational complexity of the proposed solution we derive an efficient decimation scheme by exploiting useful properties of the cascaded approach.

Research in static and time-variant system identification has brought up a broad variety of identification algorithms. In acoustics, e.g., static measurements of transfer functions are commonly conducted using Inverse Cyclic Convolution (ICC) with Exponential Sweep excitation. Identification and tracking of time-variant systems, however, often employ adaptive filter algorithms, such as the Normalized Least Mean Square (NLMS) algorithm. An interesting implementation variant is the so-called Efficient NLMS (eNLMS) algorithm for arbitrary periodic excitation. ICC and the eNLMS algorithm originate from different fields and have so far evolved independently. This paper bridges the gap using a theoretical analysis of both algorithms to prove that they can be transferred into each other. This understanding provides a joint perspective, such that know-how from both fields can be combined to further optimize the system identification process.

Occluding the ear canal alters the perception of the own voice. This is called the occlusion effect. Low frequency components of the own voice are amplified and high frequency components are attenuated. In this contribution, we extend our previously published Active Occlusion Cancellation (AOC) system for headphones and hearing aids. It uses the principle of Active Noise Cancellation (ANC) to enhance the perception of the own voice for an occluded ear canal. We combine robust feedback controller design with an adaptive mechanism to attenuate the low frequency components of the own voice. Furthermore, we include an equalizing hear-through filter to enhance high frequency components. The AOC system is evaluated in a hearing test with an in-ear headphone connected to a real-time processing system. Objective and subjective metrics show a clear improvement of the perception of the own voice. The occlusion effect is significantly reduced by the system.

In system identification one problem is the autocorrelation of the excitation signal which often crucially affects the adaptation process. This paper focuses on the Kalman filter based adaptation working in the frequency domain and the implication due to correlated signal input. Principle simulations and the introduction of a reference model indicate to which extent correlation take effect. The experimental results demonstrate that even though the Kalman approach already takes advantage from a certain level of inherent decorrelation, it also benefits from additional decorrelation. To address this issue, we derive a new realizable efficient structure combining the Kalman filter based adaptation with linear prediction techniques. The performance gains of the proposed approach are confirmed via experiments for an acoustic echo cancellation application for different scenarios.

Algorithms for Near-End Listening Enhancement (NELE) improve the intelligibility of speech from the far-end, played back in a near-end noisy environment, by adaptively filtering the speech signal and taking into account the near-end background noise characteristics. In contrast to previous works, this study considers that the speech from the far-end is also disturbed by additive noise. A noise reduction (NR) subsystem is concatenated to NELE and a joint control of NR and NELE is proposed. Listening tests confirm that the joint control leads to a significantly higher intelligibility and quality.

In announcement and communication systems, clean speech is often played back by loudspeakers in the presence of local background noise which decreases intelligibility. Speech reinforcement is a technique that enhances the intelligibility by adaptively filtering the speech, usually based on measured noise characteristics. In this contribution, solutions known from literature are applied to the case of public address systems where the microphone for noise measurement and the loudspeaker are close to each other and are far away from the listener on the ground. After establishing an acoustical model, the main problems are identified to be strong echoes between loudspeaker and microphone which make conventional noise estimation impossible and lead to an unstable system. A new approach including echo path estimation and echo-aware noise estimation is proposed and evaluated by means of simulations.

A drawback of digital transmission of analog signals is the unavoidable quantization error which leads to a limited quality even for good channel conditions. This saturation can be avoided by using analog transmission systems with discrete-time and quasi-continuous-amplitude encoding and decoding, e.g., Analog Modulo Block codes (AMB codes). The AMB code vectors are produced by multiplying a real-valued information vector with a real-valued generator matrix using a modulo arithmetic. Here, algorithms for improving the decoding performance are presented. The Lattice Maximum Likelihood (LML) decoder, a variant of the Discrete Maximum Likelihood (DML) decoder, is derived and analyzed. It refines the Zero Forcing (ZF) result if necessary, thus achieving near-ML signal quality with a reduced decoding complexity. A reduced complexity is essential for decoding high-dimensional code words. Additionally, pre- and post-processing methods are presented and analyzed, which increase the signal-to-distortion ratio (SDR) of the received symbols.

Ambisonics uses a well-respected soundfield representation and thus may become a standard in soundfield transmission, storage, and reproduction. Then we will need signal processing options in order to decode and apply the Ambisonic format universally in various applications from loudspeaker to headphone reproduction. More specifically, the use of Ambisonics is envisioned also in mobile applications where headphone reproduction is clearly preferred. In this context, we present two different system options for the advanced implementation of binaural soundfield rendering. Both options utilize the concept of virtual loudspeakers and subsequent binaural rendering via HRTF. In order to achieve spatial realism, one option relies on adaptive soundfield rotation in the Ambisonic domain to mimic a corresponding head rotation of the listener in the given soundfield. The other system option relies on a new continuous-azimuth HRTF format for direct representation of head rotations within the virtual loudspeaker setup. For both system options, we will discuss the pros and cons regarding the implementation. We conclude with a comparison of the resulting localization accuracies in listening tests.

Recommendation G.729.1 is a new ITU-T standard which was approved in May 2006. This recommendation describes a hierarchical speech and audio coding algorithm built on top of a narrowband core codec. One challenge in the codec design is the generation of a wideband signal with a very limited additional bit rate (less than 2 kb/s). In this paper, we describe the respective codec layer, which extends the transmitted acoustic bandwidth from the narrowband frequency range (50 Hz-4 kHz) to the wideband frequency range (50 Hz-7 kHz). The underlying algorithm uses a fairly coarse parametric description of the temporal and spectral energy envelopes of the high frequency band (4-7 kHz). This parameter set is quantized with a bit rate of 1.65 kb/s. At the decoder side, the high-frequency components are regenerated by appropriately shaping a synthetically generated ldquoexcitation signal.rdquo Apart from the algorithmic description and a discussion, we state a complexity evaluation as well as some listening test results.

We present a technique to extend narrowband (NB) speech communication systems, using e.g. the GSM enhanced full rate (EFR) codec, with wideband (WB, 50-7000 Hz) capability. The limited acoustic bandwidth of narrowband speech coding is extended using a fairly coarse description of the missing high frequency band (3.4-7 kHz) in terms of temporal and spectral envelopes. The high-band parameters are quantized, transmitted and then used at the receiver side to regenerate the high frequency components. The parameter encoding is done by applying split vector quantization in a transformed domain. This quantization scheme can be scaled to match any given target bit rate. Several example configurations have been implemented and tested in MUSHRA-style listening tests

This paper presents a scalable wideband speech codec working at bitrates ranging from 8 to 32 kbit/s. The core layer is the ITU-T G. 729 at 8 kbit/s. A first enhancement layer is a bandwidth extension algorithm requiring 2 kbit/s to widen the G. 729 narrow band output speech. The difference between the wideband original and reconstructed signal is transformed in the time-frequency domain by a full wavelet decomposition. The resulting coefficients are quantized by an embedded quantizer at 22 kbit/s. Listening tests show the relevance of such a scheme when compared to a pure wavelet packet decomposition. In addition, listening tests suggest that the proposed codec is equivalent to the ITU-T G. 722 at 48 kbit/s for speech signals.

We present an embedded and hierarchical 8-32 kbit/s speech and audio coding algorithm that has been successfully submitted to the ITU-T as a candidate for ITU-T Rec. G.729.1 (ex G.729EV). The submitting consortium consisted of Siemens AG, Matsushita Electric Industrial Co. Ltd., and Mindspeed Technologies Inc. This contribution gives a comprehensive overview of the proposed codec, describes the implemented algorithms, and states a detailed characterization as well as results of the official G.729EV qualification tests

In recent years research in the area of artificial bandwidth extension (BWE) of speech signals has made significant progress. The intention of BWE is to produce wideband speech with a cut-off frequency of for instance 7 kHz from a narrowband version (e.g., with telephone bandwidth, i.e., 300 Hz − 3.4 kHz). The respective algorithms are based on the estimation of parameters of a source model for speech production given the knowledge of thenarrowband signal. Atheoretical performance bound on this estimation has been formulated in [1]. In order to overcome this boundary we propose the transmission of (compact) side information which can support the parameter estimation. Since a common additional channel would conflict with the requirement of backwards compatibility in narrowband communication systems, the side information is embedded as digital watermark into the narrowband speech.

The aim of artificial bandwidth extension (BWE) is to convert speech signals with "standard telephone" quality (frequencies up to 3.4 kHz) into 7 kHz wideband speech. The principal key to high quality BWE is the estimation of the spectral envelope of the wideband speech. In general, this estimation of the wideband spectral envelope is based on a number of features that are extracted from the narrowband input speech signal. We investigate potential features and evaluate their suitability for the BWE application. The quality of each feature is quantified in terms of the statistical measures of mutual information and separability. It turns out that the best BWE results are obtained by using a large feature "super-vector" (/spl rarr/ high mutual information) which is subsequently reduced in dimension by a linear discriminant analysis (/spl rarr/ large separability). This solution also helps to reduce the computational complexity of the estimation of the wideband spectral envelope.

We present an algorithm to derive 7 kHz wideband speech from narrowband "telephone speech". A statistical approach is used that is based on a hidden Markov model (HMM) of the speech production process. A new method for the estimation of the wideband spectral envelope is proposed, using nonlinear state-specific techniques to minimize a mean square error criterion. In contrast to common memoryless estimation methods, additional information from adjacent signal frames can be exploited by utilizing the HMM. A consistent advantage of the new estimation rule is obtained compared to previously published HMM-based hard or soft classification methods.

The aim of the artificial bandwidth extension (BWE) of speech signals is to recover wideband speech from bandlimited speech. As the BWE algorithm is supposed to operate without additional side information on the original wideband speech, it has to exploit mutual dependencies between the available and missing frequency bands of the speech signal. In this paper the BWE is examined from an information theoretic perspective. After defining a performance measure, and introducing a few assumptions on a generalized BWE algorithm, a general relationship between mutual information and the maximum achievable estimation performance is formulated, which ensues an upper bound on the performance of BWE algorithms. Finally, some measurements considering a representative BWE scenario are presented.

This paper presents and compares algorithms for combined acoustic echo cancellation and noise reduction for hands-free telephones. A structure is proposed, consisting of a conventional acoustic echo canceler and a frequency domain postfilter in the sending path of the hands-free system. The postfilter applies the spectral weighting technique and attenuates both the background noise and the residual echo which remains after imperfect echo cancellation. Two weighting rules for the postfilter are discussed. The first is a conventional one, known from noise reduction, which is extended to attenuate residual echo as well as noise. The second is a psychoacoustically motivated weighting rule. Both rules are evaluated and compared by instrumental and auditive tests. They succeed about equally well in attenuating the noise and the residual echo. In listening tests, however, the psychoacoustically motivated weighting rule is mostly preferred since it leads to more natural near end speech and to less annoying residual noise.

In this contribution we optimize a speech enhancement preprocessor such that a distortion measure in the line spectral frequency (LSF) domain is minimized. We can thus improve the estimation of spectral parameters of a speech coder when the input signal to the coder is a noisy speech signal. The optimization aims at the maximum noise reduction of the enhancement preprocessor. The average maximum noise reduction characteristic is determined as a function of the speech signal SNR and is approximated by an exponential function. Since LSF parameters are widely used in speech coding the results are applicable to a wide range of speech coders and enhancement preprocessors. We report experimental results for an MMSE log spectral amplitude estimator in conjunction with the new ETSI adaptive multi-rate (AMR) speech coder. We found that the method is most effective for the low bit rate coding modes.

We address the problem of acoustic echo cancellation and noise reduction for narrow and wide band telephone applications. We combine a conventional echo canceller with a postfilter implemented in the frequency domain and derive an algorithm for the simultaneous attenuation of residual echo and noise. The main goals are a low level natural sounding background noise without artifacts such as musical tones, and an inaudible residual echo. This is achieved by considering the masking properties of the human auditory system. Simulation results verify that these goals are reached, while the distortion of the near end speech is comparable to conventional algorithms.

In this paper we propose an algorithm for reduction of noise in audio signals. In contrast to several previous approaches we do not try to achieve a complete removal of the noise, but instead our goal is to preserve a pre-defined amount of the original noise in the processed signal. This is accomplished by exploiting the masking properties of the human auditory system. The speech and noise distortions are considered separately. The spectral weighting rule, adapted by utilizing only estimates of the masking threshold and the noise power spectral density, has been designed to guarantee complete masking of distortions of the residual noise. Simulation results confirm that no audible artifacts are left in the processed signal, while speech distortions are comparable to those caused by conventional noise reduction techniques.