Gaël Mahé

Nonlinear Audio Systems Identification Through Audio Input Gaussianization

Nonlinear audio system identification generally relies on Gaussianity, whiteness and stationarity hypothesis on the input signal, although audio signals are non-Gaussian, highly correlated and non-stationary. However, since the physical behavior of nonlinear audio systems is input-dependent, they should be identified using natural audio signals (speech or music) as input, instead of artificial signals (sweeps or noise) as usually done. We propose an identification scheme that conditions audio signals to fit the desired properties for an efficient identification. The identification system consists in (1) a Gaussianization step that makes the signal near-Gaussian under a perceptual constraint; (2) a predictor filterbank that whitens the signal; (3) an orthonormalization step that enhances the statistical properties of the input vector of the last step, under a Gaussianity hypothesis; (4) an adaptive nonlinear model. The proposed scheme enhances the convergence rate of the identification and reduces the steady state identification error, compared to other schemes, for example the classical adaptive nonlinear identification.

Revisiting quantization theorem through audiowatermarking

In this paper, we propose the concept of “doping watermarking”, whose principle is to add an imperceptible noise to an host signal in order to improve its properties. Especially, our aim is to reduce the spectral support of the probability density function (PDF) of an audio signal in order to match the conditions of the quantization theorem. In this context, we develop a specific audiowatermarking algorithm and test its performance on real audio signals. This watermark allows to recover the PDF of a digital signal from a sub-quantized version of the signal, with very low error.

Perceptually controlled doping for audio source separation

The separation of an underdetermined audio mixture can be performed through sparse component analysis (SCA) that relies however on the strong hypothesis that source signals are sparse in some domain. To overcome this difficulty in the case where the original sources are available before the mixing process, the informed source separation (ISS) embeds in the mixture a watermark, which information can help a further separation. Though powerful, this technique is generally specific to a particular mixing setup and may be compromised by an additional bitrate compression stage. Thus, instead of watermarking, we propose a ‘doping’ method that makes the time-frequency representation of each source more sparse, while preserving its audio quality. This method is based on an iterative decrease of the distance between the distribution of the signal and a target sparse distribution, under a perceptual constraint. We aim to show that the proposed approach is robust to audio coding and that the use of the sparsified signals improves the source separation, in comparison with the original sources. In this work, the analysis is made only in instantaneous mixtures and focused on voice sources.

Watermark-Driven Acoustic Echo Cancellation

The performance of adaptive acoustic echo cancelers (AEC) is sensitive to the nonstationarity and correlation of speech signals. In this paper, we explore a new approach based on an adaptive AEC driven by data hidden in speech, to enhance the AEC robustness. We propose a two-stage AEC, where the first stage is a classical NLMS-based AEC driven by the far-end speech. In the signal, we embed—in an extended conception of data hiding—an imperceptible white and stationary signal, i.e., a watermark. The goal of the second stage AEC is to identify the misalignment of the first stage. It is driven by the watermark solely and takes advantage of its appropriate properties (stationary and white) to improve the robustness of the two-stage AEC to the nonstationarity and correlation of speech, and thus reduce the overall system misadjustment. We test two kinds of implementations: in the first implementation, referred to as adaptive watermark driven AEC (A-WdAEC), the watermark is a white stationary Gaussian noise. Driven by this signal, the second stage converges faster than the classical AEC and provides better performance in steady state. In the second implementation, referred to as maximum length sequences WdAEC (MLS-WdAEC), the watermark is built from MLS. Thus, the second stage performs a block identification of the first stage misalignment, given by the circular correlation watermark/preprocessed version of the first stage residual echo. The advantage of this implementation lies in its robustness against noise and undermodeling. Simulation results show the relevance of the “WdAEC” approach, compared to the classical “error-driven AEC.”

Inconsistent anticipatory postural adjustments (APAs) in rugby players: a source of injuries?

Background We are developing since 2010 with Thales and the Fédération Française de Rugby (FFR) M-Rex, a new kind of rugby scrum simulator. The study questioned whether it could improve safety and protect players from injury by using it as a tool for training/coaching the packs. Aim To explore the anticipatory postural adjustments (APAs) during the engagement of the ruck, because these predictive neck and back muscles contractions protect the spinal cord at the time of impacts, which is crucial to prevent injuries. Methods We quantified the kinematics and the EMG activities in high-level front row players during their initial engagement, when scrummaging with M-Rex. All studies were performed with one player interacting with the robot, at first, and then with the three players acting together. Results For most of the tested high-level players, the APA latencies were highly variable from trial to trial even though the engagement resulted in similar impacts. At time, the onset of the electromyography activity in the neck and back muscles showed latencies inferior to 50 ms or even close to zero prior to the impact , which rendered muscle contractions inefficient as APAs. We were also unable to identify clear muscular synergies underlying the APAs because of their great variability on a trial-to-trial basis. Finally, the APAs were not related to the amplitude of the ensuing impact and were asymmetric in most trials. All these characteristics held true, whether the player was playing alone or with two other frontline players. Conclusion Our result suggest that APAs should be systematically tested in high-level rugby players as well as in any high-level sport men at risk of neck and back injuries. Because APAs can be efficiently trained, our study paves the way to design individual position-specific injury prevention programme.

An image-inspired audio sharpness index

We propose a new non-intrusive (reference-free) objective measure of speech intelligibility that is inspired from previous works on image sharpness. We define the audio Sharpness Index (aSI) as the sensitivity of the spectrogram sparsity to the convolution of the signal with a white noise, and we calculate a closed-form formula of the aSI. Experiments with various speakers, noise and reverberation conditions show a high correlation between the aSI and the well-established Speech Transmission Index (STI), which is intrusive (full-reference). Additionally, the aSI can be used as an intelligibility or clarity criterion to drive sound enhancement algorithms. Experimental results on stereo mixtures of two sounds show that blind source separation based on aSI maximization performs well for speech and for music.

A robust signal quantization system based on error correcting codes

In this paper we propose a robust representation of a digital signal based on error correction codes. For each frame of the signal (N successive samples) a binary decomposition, as a (successive power 2) weighted sum of binary vectors, is first considered. Then, each binary vector is projected into the set of codewords of a corresponding block code. The codes are designed so that their correction powers increases inversely to the weight of the binary vectors since the binary vectors with high weight are less sensitive to disturbance. The corresponding representation (decoding) thus appears as a form of signal quantization that can provide an interesting protection against noise and/or channel distorsion. Some applications showing the utility of the proposed representation are given.

High-frequency tonal components restoration in low-bitrate audio coding using multiple spectral translations

At reduced bitrates, the audio compression affects high frequency tonal components of signals, which results in a roughness phenomenon. Audio coders are limited in the reconstruction of the high-frequency spectrum mainly because of the potential unpredictability of the structure of the latter, as well as unprecise indicators of tonal to noise ratio. We propose a technique for high-frequency tones restoration, based on the correction of the tonal positions in the decoded signal, using a small set of information transmitted through an auxiliary channel at a very low bit-rate (typically <; 2 kbps). The proposed approach is evaluated using objective measures of perceptual roughness. The experimental results with HE-AAC coding at 16 kbps exhibits an efficient preservation of the harmonicity and a significant improvement of the audio quality.

Restoration of the voice timbre in telephone networks based on both voice and line properties

The voice timbre suffers from different forms of distortions in a telephone link. In this paper, we propose a new blind equalizer for correcting these distortions, based on the combination of two different methods. The first one is a blind equalization method consisting in matching the long term spectrum of the processed signal to a reference spectrum, while the second one is a precompensation method, based on the physical characteristics of transmission lines. The new method is compared to the first one, showing a significant gain in performance.