Traitement parametrique des signaux audio dans le contexte des protheses auditives.

摘要

The digital hearing aids for the hearing-impaired constitute a particularly important application field of signal processing. The last generations of hearing aids are equipped with a wireless local area network which allows a sophisticated binaural processing of the exchanged information not only to improve its intelligibility and its quality perception but also to emphasize the associated source localization acoustic cues. This technological progress has made possible the use of conventional parametric processing techniques which are widely used for speech compression in telephony.;Within the context of the first approach, we study an adaptive filtering method for noise reduction, proposed previously in the literature, and we identify some inherent problems associated with its use. To fix these problems, we suggest modifying an existing noise power spectrum estimator and combining it with this filtering method by means of a soft-decision scheme. We show that this combination is advantageous in that it provides a good tradeoff between the amount of noise reduction and the speech distortion.;Regarding the second approach, we propose two versions of an iterative method for parametric processing based on an autoregressive linear model, a predefined decision criterion and the Minimum Statistics (MS) algorithm. We show that this method, which operates in the correlation domain, allows obtaining compensated and stable spectral parameters. We present the limits associated with the use of the MS algorithm in the correlation domain. Then, a method that allows improving the precision of the noise variance estimator in the case of an audio signal contaminated by an additive white noise was developed. In comparison with the method based on the MS algorithm, we show empirically that the new proposed method gives rise to more robust spectral parameters which suitably model the formantic structure of the signal.;The sensitivity problem of the conventional parametric processing techniques to background noise is well-known. Two approaches of parametric processing in the presence of noise are often used. The first approach consists of performing the reduction of noise, generally in the frequency domain, as a preprocessing for the parametric modeling of the audio signal. The second approach allows applying a noise reduction processing directly in the correlation domain. This processing allows estimating the spectral parameters which model suitably the formantic structure of the signal. We propose innovative and relevant algorithmic solutions to determine which of both approaches allows obtaining the best fidelity of the spectral envelope of an audio signal to its formantic structure.