IEEE ICASSP 2016 - International Conference on Acoustics, Speech and Signal Processing, Shanghai, Chine, 20-/03/2016 - 25/03/2016International audienceDeep Neural Networks (DNNs) have recently become a popular technique for regression and classification problems. Their capacity to learn high-order correlations between input and output data proves to be very powerful for automatic speech recognition. In this paper we investigate the use of DNNs for automatic scream and shouted speech detection, within the framework of surveillance systems in public transportation. We recorded a database of sounds occurring in subway trains in real conditions of exploitation and used DNNs to classify the sounds into screams, shouts and other categories. We report encouraging results, given the difficulty of the task, especially when a high level of surrounding noise is present.Les réseaux de neurones profonds sont devenues récemment une technique populaire pour les problèmes de régression et de classification. Leur capacité d'apprendre des corrélations d'ordre éleÎ entre des entrées et des données de sortie s'aÏre être très un puissant outil pour reconnaissance automatique de la parole. Dans cet article, nous étudions l'utilisation des réseaux de neurones profonds pour la détection automatique de cris et de parole criée dans le cadre de systèmes de surveillance dans les transports publics. Pour cela, une base de données sonores a été enregistrée dans une rame de métro en condition réelle d'exploitation. Dans ce contexte, la détection de cri est réalisée via un classement de divers types de production de la parole dont des cris. Nous obtenons des résultats encourageants étant donné la difficulté de la tâche, en particulier vis-à-vis du haut niveau de bruit sonore environnant

In this paper, a novel design scheme for radar signal is proposed. It is based on a modification of Costas codes design obtained by increasing the frequency separation of frequency hops beyond the orthogonality condition. This condition states that frequency separation between frequency hops should be at most equal to the inverse of bit duration of the code. Indeed, increasing frequency separation beyond this condition is sometimes needed in radar applications. However, this leads to the appearance of grating lobes near the main temporal lobe of the autocorrelation function. This paper proposes overcoming this problem using sub-pulse coding. The originality of the proposed approach consists in choosing for Costas sub-pulses other Costas codes instead of constant frequency ones in order to lower autocorrelation grating lobes that would normally appear when the orthogonality condition is not satisfied. Indeed, we show that, under some conditions on Costas signal parameters in the sub-pulse, good elimination of grating lobes is achieved.

Costas codes are a variant of pulse compression waveforms, largely studied for their attractive time-frequency properties. Their 'thumbtack-like' ambiguity function (AF) makes them highly suitable for delay and Doppler estimation, in radar and sonar applications. However, this behaviour depends heavily on the length of the code: the improvement in delay-Doppler resolutions and AF sidelobes level needs an increase in the size of the code. In this study, designs that allow good performance without increasing the size of the code are proposed. They are based on a modification of Costas codes by widening frequency separation between hops and replacing rectangular pulses by other waveforms. This will lead to a removal of autocorrelation function grating lobes that normally appear when frequency separation is increased. The originality of the work lies in the proposal of diversified pulse waveforms, such as phase codes, Slepian sequences, and other Costas codes, to encode main Costas pulses. A performance comparison of the proposed approaches is supplied. Such waveforms could also be of interest for applications where waveform diversity is desired

No abstract

No abstract