No description available.

No description available.

Human-like performance by machines in tasks of speech and audio processing has remained an elusive goal. In an attempt to bridge the gap in performance between humans and machines there has been an increased effort to study and model physiological processes. However, the widespread use of biologically inspired features proposed in the past has been hampered mainly by either the lack of robustness across a range of signal-to-noise ratios or the formidable computational costs. In physiological systems, sensor processing occurs in several stages. It is likely the case that signal features and biological processing techniques evolved together and are complementary or well matched. It is precisely for this reason that modeling the feature extraction processes should go hand in hand with modeling of the processes that use these features. This research presents a front-end feature extraction method for audio signals inspired by the human peripheral auditory system. New developments in the field of machine learning are leveraged to build classifiers to maximize the performance gains afforded by these features. The structure of the classification system is similar to what might be expected in physiological processing. Further, the feature extraction and classification algorithms can be efficiently implemented using the low-power cooperative analog-digital signal processing platform. The usefulness of the features is demonstrated for tasks of audio classification, speech versus non-speech discrimination, and speech recognition. The low-power nature of the classification system makes it ideal for use in applications such as hearing aids, hand-held devices, and surveillance through acoustic scene monitoring

A hearing impairment is primarily diagnosed by measuring the hearing thresholds at a range of auditory frequencies (air-conduction audiometry). Although this clinical procedure is simple, affordable, reliable and fast, it does not offer differential information about origins of the hearing impairment. The main goal of this thesis is to quantitatively link specific cochlear pathologies to certain changes in the spectral and temporal characteristics of the auditory system. This can help better understand the underlying mechanisms associated with sensorineural hearing impairments, beyond what is shown in the audiogram. Here, an electromechanical signal-transmission model is devised in MATLAB where the parameters of the model convey biological interpretations of mammalian cochlear structures. The model is exploited to simulate the cell-level cochlear pathologies associated with two common types of sensorineural hearing impairments, 1: presbyacusis (age-related hearing impairment) and, 2: noise-induced hearing impairment. Furthermore, a clinical study, consisting of different psychoacoustic and physiological tests, was performed to trace and validate the model predictions in human. The results of the clinical tests were collated and compared with the model predictions, showing a reasonable agreement. In summary, the present model provides a biophysical foundation for simulating the effect of specific cellular lesions, due to different inner-ear diseases and external insults, on the entire cochlear mechanism and thereby on the whole auditory system. This is a multidisciplinary work in the sense that it connects the ‘biological processes’ with ‘acoustic modelling’ and ‘clinical audiology’ in a translational context.

L'efficacité de la reconnaissance auditive peut être décrite et quantifiée suivant deux aspects différents : la quantité d'information nécessaire pour y parvenir et sa rapidité. L'objectif de cette thèse est d'évaluer expérimentalement ces deux aspects. Dans une première partie expérimentale, nous nous sommes intéressés à la quantité d'information en créant des représentations parcimonieuses de sons naturels originaux appelées esquisses auditives. Nous avons montré qu'une esquisse auditive est reconnue malgré la quantité très limitée d'information auditive présente dans les stimuli. Pour l'analyse des stimuli auditifs, nous avons développé un modèle de distance auditive entre catégories sonores. Pour l'analyse des performances des participants, nous avons développé un modèle pour le calcul de la sensibilité par catégorie sonore et tenant compte du biais, qui s'intègre dans la théorie de détection du signal. Ces analyses nous ont permis de montrer qu'en réalité les résultats ne sont pas équivalents entre les différentes catégories sonores. La voix se démarque des autres catégories testées (e.g. instruments de musique) : la technique de sélection de l'information parcimonieuse ne semble pas adaptée aux indices de la voix. Dans une seconde partie expérimentale, nous avons étudié le décours temporel de la reconnaissance auditive. Afin d'estimer le temps nécessaire au système auditif pour reconnaître un son, nous avons utilisé un récent paradigme de présentation audio séquentielle rapide (RASP, pour Rapid Audio Sequential Presentation). Nous avons montré que moins de 50 ms suffisent pour reconnaître un son naturel court, avec une meilleure reconnaissance pour la pour la voix humaine. / The efficacy of auditory recognition relies on two different aspects: the quantity of information necessary and the processing speed. The objective of this thesis was to experimentally evaluate these two aspects. In a first experimental part, we explored the amount of information by creating sparse representations of original natural sounds to form what is called auditory sketches. We showed that an auditory sketch is recognizable despite the very limited quantity of auditory information in the stimuli. To achieve these results, we dedicated an important part of our work on the elaboration of adequate tools in function of the tested sound categories. Thus, for the analysis of auditory stimuli, we have developed an auditory distance model between sound categories. For the analysis of the performances of the participants, we have developed a model to calculate the sensitivity by sound category and taking into account the bias, which falls within the signal detection theory. These analyses allowed us to show that, actually, the results are not equivalent between the different sound categories. Voices stand out from the other categories tested (e.g. musical instruments): the technique of selection of the sparse information does not seem adapted to the voice features. In a second experimental part, we investigated the temporal course of auditory recognition. To estimate the time necessary for the auditory system to recognize a sound, we used a recent paradigm of Rapid Audio Sequential Presentation (RASP). We showed that less than 50 ms are enough to recognize a short natural sound, with a better recognition for the human voice.