In non-linear autoregressive models, the time dependency of coefficients is often driven by a particular time-series which is not given and thus has to be estimated from the data. To allow model evaluation on a validation set, we describe a parametric approach for such driver estimation. After estimating the driver as a weighted sum of potential drivers, we use it in a non-linear autoregressive model with a polynomial parametrization. Using gradient descent, we optimize the linear filter extracting the driver, outperforming a typical grid-search on predefined filters.

In this paper, we introduce new parametric generative driven auto-regressive (DAR) models. DAR models provide a nonlinear and non-stationary spectral estimation of a signal, conditionally to another exogenous signal. We detail how inference can be done efficiently while guaranteeing model stability. We show how model comparison and hyper-parameter selection can be done using likelihood estimates. We also point out the limits of DAR models when the exogenous signal contains too high frequencies. Finally, we illustrate how DAR models can be applied on neuro-physiologic signals to characterize phase-amplitude coupling.

We address the issue of reliably detecting and quantifying cross-frequency coupling (CFC) in neural time series. Based on non-linear auto-regressive models, the proposed method provides a generative and parametric model of the time-varying spectral content of the signals. As this method models the entire spectrum simultaneously, it avoids the pitfalls related to incorrect filtering or the use of the Hilbert transform on wide-band signals. As the model is probabilistic, it also provides a score of the model "goodness of fit" via the likelihood, enabling easy and legitimate model selection and parameter comparison; this data-driven feature is unique to our model-based approach. Using three datasets obtained with invasive neurophysiological recordings in humans and rodents, we demonstrate that these models are able to replicate previous results obtained with other metrics, but also reveal new insights such as the influence of the amplitude of the slow oscillation. Using simulations, we demonstrate that our parametric method can reveal neural couplings with shorter signals than non-parametric methods. We also show how the likelihood can be used to find optimal filtering parameters, suggesting new properties on the spectrum of the driving signal, but also to estimate the optimal delay between the coupled signals, enabling a directionality estimation in the coupling.

Reverberation degrades speech intelligibility in telecommunications as well as it increases the word error rate in automatic speech recognition tasks. Several dereverberation methods have been proposed recently in order to counter these effects. In the single microphone case, the dereverberation problem is underdetermined and reverberation suppression approaches are preferred. In this paper we propose a novel method for single channel reverberation suppression. Late reverberation is estimated in the time-frequency domain as a sparse linear combination of previous frames. The predictors associated to the model are determined in a Lasso framework and a spectral subtraction filter is designed to produce the enhanced signal. This model does not require any additional information about the room acoustics and it is well suited for real-time applications. The method has state-of-the-art performance in terms of both reverberation suppression and spectral distortion.

We present a two stage blind source separation (BSS) algorithm for robot audition. The algorithm is based on a beamforming preprocessing and a BSS algorithm using a sparsity separation criterion. Before the BSS step, we filter the sensors outputs by beamforming filters to reduce the reverberation and the environmental noise. As we are in a robot audition context, the manifold of the sensor array in this case is hard to model, so we use pre-measured Head Related Transfer Functions (HRTFs) to estimate the beamforming filters. In this article, we show the good performance of this method as compared to a single stage BSS only method.

This work addresses the Huawei/3Dlife Grand challenge proposing a set of audio tools for a virtual dance-teaching assistant. These tools are meant to help the dance student develop a sense of rhythm to correctly synchronize his/her movements and steps to the musical timing of the choreographies to be executed. They consist of three main components, namely a music (beat) analysis module, a source separation and remastering module and a dance step segmentation module. These components enable to create augmented tutorial videos highlighting the rhythmic information using, for instance, a synthetic dance teacher voice, but also videos highlighting the steps executed by a student to help in the evaluation of his/her performance.

This paper describes a probabilistic approach to template-based chord recognition in music signals. The algorithm only takes chromagram data and a user-defined dictionary of chord templates as input data. No training or musical information such as key, rhythm, or chord transition models is required. The chord occurrences are treated as probabilistic events, whose probabilities are learned from the song using an expectation-maximization (EM) algorithm. The adaptative estimation of these probabilities (together with an ad-hoc postprocessing filtering) has the desirable effect of smoothing out spurious chords that would occur in our previous baseline work. Our algorithm is compared to various methods that entered the Music Information Retrieval Evaluation eXchange (MIREX) in 2008 and 2009, using a diverse set of evaluation metrics, some of which are new. The systems are tested on two evaluation corpuses; the first one is composed of the Beatles catalog (180 pop-rock songs) and the other one is constituted of 20 songs from various artists and music genres. Results show that our method outperforms state-of-the-art chord recognition systems.

In this paper, we propose a simple and fast method for chord recognition in music signals. We extract a chromagram from the signal which transcribes the harmonic content of the piece over time. We introduce a set of chord templates taking into account one or more harmonics of the pitch notes of the chord and calculate a scale parameter to fit the chromagram frames to these chords templates. Several chord types (major, minor, dominant seventh, etc.) are considered. The detected chord over a frame is the one minimizing a measure of fit between the rescaled chroma vector and the chord templates. Several popular distances and divergences from the signal processing or probability fields are considered for our task. Our system is improved by some post-processing filtering that modifies the recognition criteria so as to favor time-persistence. The transcription tool is evaluated on three corpora: the Beatles corpus used for MIREX 08, a 20-audio-song corpus, and a resynthesized MIDI corpus. Our system is also compared to state-of-the-art chord recognition methods. Experimental results show that our method compares favorably to the state-of-the-art and is less computationally demanding than the other evaluated systems. Our systems entered the MIREX 2009 competition and performed very well.

This paper describes a method for chord recognition from audio signals. Our method provides a coherent and relevant probabilistic framework for template-based transcription. The only information needed for the transcription is the definition of the chords : in particular neither annotated audio data nor music theory knowledge is required. We extract from the signal a succession of chroma vectors which are our model observations. We propose a generative model for these observations from chord distribution probabilities and fixed chord templates. The parameters are evaluated through an EM algorithm. In order to capture the temporal structure, we apply some post-processing filtering methods before detecting the chords. Our method is evaluated on two audio corpus. Results show that our method outperforms state-of-the-art chord recognition methods and also gives more relevant chord transcriptions.

This paper considers the blind separation of nonstationary sources in the underdetermined convolutive mixture case. We introduce two methods based on the sparsity assumption of the sources in the time-frequency (TF) domain. The first one assumes that the sources are disjoint in the TF domain; i.e. there is at most one source signal present at a given point in the TF domain. In the second method, we relax this assumption by allowing the sources to be TF-nondisjoint to a certain extent. In particular, the number of sources present (active) at a TF point should be strictly less than the number of sensors. In that case, the separation can be achieved thanks to subspace projection which allows us to identify the active sources and to estimate their corresponding time-frequency distribution (TFD) values.

This paper considers the blind separation of nonstationary sources in the underdetermined convolutive mixture case. We introduce, two methods based on the sparsity assumption of the sources in the time-frequency (TF) domain. The first one assumes that the sources are disjoint in the TF domain, i.e., there is at most one source signal present at a given point in the TF domain. In the second method, we relax this assumption by allowing the sources to be TF-nondisjoint to a certain extent. In particular, the number of sources present (active) at a TF point should be strictly less than the number of sensors. In that case, the separation can be achieved thanks to subspace projection which allows us to identify the active sources and to estimate their corresponding time-frequency distribution (TFD) values. Another contribution of this paper is a new estimation procedure for the mixing channel in the underdetermined case. Finally, numerical performance evaluations and comparisons of the proposed methods are provided highlighting their effectiveness.

Sound field analysis is performed from data recorded by a microphone array in order to extract some spatial information about the sound field. In this article, the assets of the multidimensional Fourier transform are emphasized in order to perform this analysis step. Indeed, the finite number of microphones inevitably introduces spatial aliasing, which is linked to the spectral broadening due to the discrete and finite nature of the analysis window. First, given the geometry of the array, the hypothesis that sound fields satisfies the dispersion relation, obtained from the wave equation, is used to design an optimal analysis window, which minimizes spatial aliasing according to a criterion inspired by prolate spheroidal spectral analysis. Then, several array configurations are compared to analyze sound fields with large frequency bandwidth

A fast iterative algorithm, with computation based on the fast Fourier transform (FFT), is presented. It can be used to control a sound field at several control points with a loudspeaker array from multiple reference signals. It designs an equalizer able to invert long FIR filters and which achieves better performance than traditional FFT-based deconvolution methods with an equal number of coefficients in the inverse filters.

This paper deals with the estimation of fundamental frequency for two speakers recorded on the same channel. We estimate a set of speech models based upon sinusoidal plus autoregressive noise representations. We then detect the best model from this set using Rissanen criterion. Equivalent to a penalized log-likelihood, the criterion is also used to carry out a voicing detection. The detector compares the likelihood of a sinusoid plus noise model with the likelihood of a simple autoregressive model. Several simulations are presented to illustrate this estimation method.

Adaptive acoustic echo cancellation in stereophonic teleconferencing is a very demanding application. The characteristics are: a very large number of coefficients, non-stationary input (speech), (slowly) time-varying systems to be identified, plus the specific property that both stereo signals are intrinsically very correlated. Basic versions of stochastic gradient algorithms have difficulties in meeting these requirements. We show that, in a multichannel framework, only a combination of techniques can result in an algorithm which convergence is governed by a quasi-diagonal matrix. Simulations with data recorded in a conference room demonstrate the improvement in convergence of our algorithm compared to the LMS.

Matched filtering and adaptive beamforming are both necessary for efficient speech dereverberation and noise reduction by microphone arrays. This can be achieved by the identification of impulse responses. We show that adaptive microphone arrays are sensitive to the identification errors of the impulse responses, particularly due to speaker movements. We prove that adjusted matched-filtering and permanent tracking of the impulse responses are also necessary. The proposed microphone array responds well to these requirements under realistic conditions.

This paper presents a method of adaptive microphone array beamforming using matched filters with signal subspace tracking. Our objective is to enhance near-field speech signals by reducing multipath and reverberation. In real applications such as speech acquisition in acoustic environments, sources do not propagate along known and direct paths. Particularly in hands-free telephony, we have to deal with undesired propagation phenomena such as reflections and reverberation. Prior methods developed adaptive microphone arrays for noise reduction after a time delay compensation of the direct path. This simple synchronization is insufficient to produce an acceptable speech quality, and makes adaptive beamforming unsuitable. We prove the identification of source-to-array impulse responses to be possible by subspace tracking. We consequently show the advantage of treating synchronization as a matched filtering step. Speech quality is indeed enhanced at the output by the suppression of reflections and reverberation (i.e., dereverberation), and efficient adaptive beamforming for noise reduction is applied without risk of signal cancellation. Evaluations confirm the performance achieved by the proposed algorithm under real conditions.

We propose in this paper an array of microphones which achieves good performance in double talk situations. By a subspace tracking procedure, we jointly identify the acoustic paths from the speech and echo sources. With an adaptive beamformer constrained over these paths, we properly recover the original speech, cancel the echo and reduce the background noise. Simulations made with real data confirm the efficiency of this array.

We propose a new fast projection algorithm for stereophonic acoustic echo cancellation. This algorithm can be viewed as a generalization of the extended two-channel LMS algorithm which takes into account the correlation between the input signals. Moreover, this algorithm fits naturally with the framework of affine projection techniques extended to the two-channel case. Its computational complexity is less than half the complexity of the fastest two-channel RLS versions. Simulation results show the obtained performance.

We consider the problem of locating and extracting sources received by an array of sensors in two significant cases. In one case, sensor positions are partially unknown. In the other one, the propagation to a subset of sensors is unknown. The core of the algorithm is partially blind identification of steering vectors. It relies on a tracking procedure with low computational complexity. We apply it to the calibration of maneuvering and towed arrays receiving multiple and correlated sources.

A straightforward generalization of the so-called affine projection algorithm (APA) to the multichannel (MC) case is easily obtained. However, due to the strong correlation between the input signals of the various channels, the resulting algorithm converges very slowly. The article describes the way to overcome this problem and derives an efficient algorithm that makes use of additional orthogonal projections.

It is likely that stereophonic (and more generally, multichannel) sound pick-up, transmission and diffusion will be implemented in future teleconference systems to provide the users with enhanced quality. Therefore, adequate solutions must be found to solve the problem of stereophonic acoustic echo which will occur in such systems. We explain in this paper the difference between the mono and two-channel systems and the behavior of the two-channel classical adaptive algorithms in comparison with the same algorithms in the mono-channel case. Also, we outline a new NLMS-like algorithm derived from the two-channel RLS algorithm as a first member of a family of improved two-channel adaptive filters.

We present a novel adaptive algorithm in the frequency domain with a low order of arithmetic complexity for simultaneously performing beamforming, source direction finding and array shape calibration. The algorithm is proposed for multiple wideband sources, but could be applied to the narrowband case or to a single wideband source. The source signals are first estimated using a set of beamformers. These estimates are processed with the observation signals to track the steering vectors within the signal subspace. The adapted steering vectors are projected over the array manifold, to finally estimate the source directions and sensor positions. Simulations show the efficiency of the algorithm to achieve the proposed tasks.

Sensor positioning has an important influence on the performance of array processing. The main contributions of this paper are criteria to evaluate and to optimize the performance of an array of sensors through its geometric configuration-the number and the positions of sensors. The distance between the observed and the desired directivity patterns has been used as a criterion. We propose a generalization of this criterion. We obtain a second criterion by an orthogonality measurement between two subspaces, respectively, containing the noise and the desired signals. This measurement is obtained by an integration of the variance of the signal estimated by an unbiased linear estimator. The output power of a minimum variance distortionless response (MVDR) filter is considered as another criterion. Considering the channels between the desired sources and the array as a multidimensional communication channel, we obtain as a criterion the channel information capacity, which shows the maximum information that we may transfer by this channel. The simulation results confirm that the criteria are related and properly characterize the array configuration. For a linear equally spaced array and narrowband plane-wave signals, the classical results regarding the sensor spacing are confirmed.>

This paper presents an algorithm implementing robustness in beamforming, by directly reducing localization errors in the presence of pointing errors or a single moving target. Given an initial position, the desired source signal is first estimated using a classical beamforming unit. It is in a second step, processed by an LMS-like gradient stochastic estimation procedure of the steering vector, to adaptively track the correct source position. The newly identified source position is projected over the array manifold, then finally transmitted in a feedback loop to the beamforming unit, closing in this way the global algorithm iteration. The simulation results show that robustness is effectively realized without any compromising output SNR loss. Moreover, they prove an efficient tracking behavior in the presence of mobile sources. >

Sensor positioning has an important influence on the performance of array processing. The main contributions of the paper are the proposal of criteria to evaluate and to optimize the performance of an array of sensors through its geometric configuration, the number and the positions of the sensors. Each criterion evaluates the performance of the array for some applications, particularly for beamforming and source localization. The simulation results show that the criteria are related, and allow proper optimization of the array configuration.>

The authors deal with frequency-domain (FD) adaptive identification with application to acoustic echo cancellation. Several proposed frequency-domain LMS algorithms are reviewed and unified in the common framework of fast convolution using the short-term Fourier transform. A more flexible identification scheme is then proposed, making it possible to independently control the delay and the adaptation rate. The theoretical proof of convergence of this new scheme is provided together with simulations demonstrating the efficiency of this method in both stationary and nonstationary environments.>

The authors describe a microphone array for speech recording in car environments. The array is designed for hands-free radiotelephone, and is also used as a front-end for an automatic speech recognition system. The configuration of the array and the adaptive beamforming technique implemented are described and the performance of the array are evaluated. The measure of performance is the score obtained with a speech recognition system. Within the European ESPRIT project ARS (adverse environment recognition of speech), a prototype of this microphone array was built. It has eight microphones and works in real time, using one TMS C30 processor.>

A real-time speech coder which operates at 9.6 kb/s is described. In order to favor the excitation coding, an original structure called the vector excited adaptive predictive coder (VEAPC) is presented where neither the short-term predictor nor the long-term predictor is quantized and transmitted to the receiver. According to scalar ADPCM techniques, this approach uses backward adaptive predictive coding to sequentially update the short-term predictor coefficients. The ADPCM scheme is generalized to the vector case as long as the quantized and transmitted error is a vector. The excitation vector computation leads to an analysis by synthesis procedure where fast CELPC/VXC algorithms are efficiently used.>

A concept of model expansion that generalizes several methods already present in some time-varying autoregressive techniques is presented. These methods involve a representation of the predictors in terms of a set of mathematical functions of time. This has two effects: replacing a scalar problem by a vector one, (which is a negative effect since it increases the dimension of the model), and replacing a nonstationary problem by a time-invariant one (which is a positive effect, since it allows estimation through time averages). The generalization is an embedding technique that simplifies a nonstationary or nonlinear problem to a linear invariant problem by replacing the original signals by a larger set of signals. Three applications of this technique are given: analysis of nonstationary speech signals, location of broadband sources received on an array of equispaced sensors, and location of narrowband sources in the near field of a linear antenna.>

A linear system representation called the double convolution model, which combines the impulse response and transfer function, is presented. It is based on Fourier transformation by block of the impulse response. Its main advantage is that the model order and the window size are not imposed by the response length. This property makes the adaptation of systems with long impulse response easier because it may be possible to use a moderate order and to avoid large processing delay. Simulation results are presented for an acoustic channel, the impulse response of which is measured in a reverberant room.>

The authors present a new model to characterize the trajectories of the analysis parameters of speech. They represent their evolution as the output of a vector autoregressive system excited by a multistep function, i.e., a sum of an arbitrary number of step-functions. The constant parts of the resulting function represent targets to be attained by the output of the system. Each target corresponds roughly to a speech event. As a new target can arrive before the present one is attained, coarticulation phenomenon can be appropriately treated. Inversely, stable parts of speech are represented by sufficiently spaced targets. Detection of steps and estimated of their amplitudes are performed iteratively by an analysis-by-synthesis method that minimises a quadratic criterion.>

This paper describes a technique of spectral transformation for improved adaptation of a knowledge data base or reference templates to new speakers in automatic speech recognition (ASR). Based on a statistical analysis tool (Canonical correlation analysis) the proposed method permits to improve speaker independance in Large vocabulary ASR. Application to an isolated word recognizer improved a 70% correct score to 87%.

A large class of nonstationary signals is efficiently modeled through time-varying autoregressive or ARMA models which are characterized by the constraint that their coefficients evolve within a subspace, a basis of functions spanning it being known prior to estimation. In this context, powerful estimators exist for AR models, and lattice filters, but for the more complex ARMA models, existing estimators are either too costly or too poor. This paper presents a new estimator which realizes a good compromise between cost and quality. The AR and MA part of the model are estimated simultaneously in order to fit the two-indices nonstationary impulse response of a long time-varying autoregressive model. In the case where the AR order is equal to the MA order plus one, the solution has a Toëplitz structure and leads to a fast algorithm.

In natural speech production, the vocal tract continuously changes in shade and therefore operates as a time-varying Filter. In this paper, the time-dependent (autoregressive, PARCOR or Log Area Ratio) coefficients of an AR model are appoximated by a linear decomposition on the basis of orthogonal (Legendre, Fourier and orolate spheroidal) functions of time, covering an entire speech segment (oF syllabic size). These techniques are validated by synthesis experiments. Listening tests have demonstrated that spectral encodinq using non-stationary models allows an average 50% reduction of the number of parameters to be transmitted per second of speech, as compared to spectral estimators operating on a frame by frame basis. Applications to speech coding, synthesis and recognition are discussed.

A large class of non-stationary signals, containing speech signals, but not restricted to them, can be represented by time-varying models, the coefficients of which are finite linear combinations of known time functions. Such models have been found useful for speech recognition and speech synthesis but they suffer in this last application from a lack of stability. This paper describes a time-varying AR model into which the time-dependency is coded through Log Area Ratios: this ensures its stability. Two algorithms for the estimation of these time-varying Log Area Ratios are proposed, the first one is an approximation via a lattice filter, while the second one maximises a least squares criterion.

This paper discusses the estimation of time-frequency representations of non-stationary signals, by means of ARMA (autoregressive moving-average) models with time-dependent coefficients. In order to allow the estimation of the model on a single realisation of the random signal, it is assumed that the coefficients are weighted sums of known functions, for instance polynomials, cosines. Two ARMA models will be discussed in this paper. The first one is a white noise driven ARMA model, which is rather usual. The second one has been designed for deterministic or speech-like signals and consists of a linear system driven by intermittent inputs with additive output (white) noise. The time-frequency representation of the signal is then obtained from the state-space realisation of the time-dependent model. After a description of the algorithms involved in both estimators, the paper focuses on a comparison of the two ARMA models on modulated sinusoids corrupted by white noise. This shows that these models behave equally well for high SNR (signal-to-noise ratio), but the deterministic model gives better results when the SNR decreases.

This paper describes two methods for time-dependent moving-average modelling of non-stationary signals. In both of them, the finite-order MA model is approximated by an infinite-order auto-re gressive model. In the second one, the use of the AR model is explicit : by means of the inversion of this model, the innovation is estimated and the MA model is then obtained by a least-square procedure. In the first method, the AR model is implicit : a vector signal is associated to the scalar one, and the MA algorithm contains two steps : Schur parametrization of an estimate of the covariance of the vector process, followed by a reduction of the vector model to a scalar one.

Modeling of non-stationary signals can be achieved through time-dependent lattices and auto-regressive models. This method leads to an extension of several well-known techniques of stationary spectral estimation to the non-stationary case. Time varying AR models are identified by means of a fast Levinson algorithm, while the time-dependent reflexion coefficients of the lattice are estimated through an algorithm which is similar to Burg's. The context of these methods is off-line (blockwise) estimation.

Linear prediction is a well extended technique for transmission, synthesis and recognition. However when the signal is corrupted by noise, the estimation of the auto-regressive model is known to be biaised. This paper is devoted to methods allowing a reduction of this bias. We will consider first a global method, in which the Yule Walker equations are modified to take into account the variance of an additive white noise. The problem becomes non-linear and is solved recursively. In a second approach, we will examine a time - recursive method based on Kalman filtering.

A procedure for the adaptation of a phonetic recognition system to a new speaker, is described. The reference for each phoneme(cepstrum of the autoregressive model) is transformed linearly in order to fit optimally the utterances of a new speaker. In a first step, samples uttered by the new speaker are mapped onto corresponding utterances (reference) through a dynamic comparison. In a second step, the linear transformation is computed through canonical correlation analysis of the samples. Experimental simulations provided satisfactory results.

The classification of phonemes, issued from a distance measure based on the linear prediction residuals is investigated, focusing on the dependency of the scores (of good recognition) upon the identity of the speaker. Starting from templates well adapted to the voice of one (standard) speaker, a scheme is proposed in order to improve the templates accordingly to global properties of the voice of the new speaker. The required transformation is linear and obtained by means of the statistical tool of canonical correlation analysis. Experiments carried out over six speakers show the adequacy of this transformation.

The paper emphasizes the importance of normalization of parameters in identification of linear models as now commonly applied to digital signal processing. Classical LPC, Pisarenko, Prony methods are unified and compared. The factorial approach plays a central role when additive noise is considered. The computational requirement is the determination of eigen vectors of correlation and covariance matrices. Various algorithms are then given including sequential estimation procedures in the covariance case. The methods are compared on close sinewaves merged in noise in terms of resolution, windowing, signal-to-noise ratio.

Using the formulation of Levinson's algorithm in terms of cross correlation estimates as presented by RISSANEN, it is shown that, in the case of a finite length autocorrelation serieR(z)(i.e. moving average models), this procedure generates a serie that converges towards a causal impulse responseE(z). This response has a finite length and its autocorrelationE(z)\cdotE(z^{-1})is proportionnal toR(z). Moreover it is shown thatE(z)has all its roots inside or on the unit cercle (minimal phase response). The proof uses mainly the properties of the PARCOR coefficients.