Recent research on machine learning focuses on audio source identification in complex environments. They rely on extracting features from audio signals and use machine learning techniques to model the sound classes. However, such techniques are often not optimized for a real-time implementation and in multi-source conditions. We propose a new real-time audio single-source classification method based on a dictionary of sound models (that can be extended to a multi-source setting). The sound spectrums are modeled with mixture models and form a dictionary. The classification is based on a comparison with all the elements of the dictionary by computing likelihoods and the best match is used as a result. We found that this technique outperforms classic methods within a temporal horizon of 0.5s per decision (achieved 6% of errors on a database composed of 50 classes). Future works will focus on the multi-sources classification and reduce the computational load.

This work addresses the recurring challenge of real-time monophonic and polyphonic audio source classification. The whole normalized power spectrum (NPS) is directly involved in the proposed process, avoiding complex and hazardous traditional feature extraction. It is also a natural candidate for polyphonic events thanks to its additive property in such cases. The classification task is performed through a nonparametric kernel-based generative modeling of the power spectrum. Advantage of this model is twofold: it is almost hypothesis free and it allows to straightforwardly obtain the maximum a posteriori classification rule of online signals. Moreover it makes use of the monophonic dataset to build the polyphonic one. Then, to reach the real-time target, the complexity of the method can be tuned by using a standard hierarchical clustering preprocessing of the prototypes, revealing a particularly efficient computation time and classification accuracy trade-off. The proposed method, called RARE (for Real-time Audio Recognition Engine) reveals encouraging results both in monophonic and polyphonic classification tasks on benchmark and owned datasets, including also the targeted real-time situation. In particular, this method benefits from several advantages compared to the state-of-the-art methods including a reduced training time, no feature extraction, the ability to control the computation -accuracy trade-off and no training on already mixed sounds for polyphonic classification.

Binaural cues such as the interaural time and level differences are the primary cues for estimation of the lateral position of a sound source, but are not sufficient to determine elevation or the exact position on the "cone of confusion". Spectral content of the head-related transfer function (HRTF) provides cues that permit this discrimination, notably high frequency peaks and notches created by diffraction effects within the pinnae. For high quality binaural synthesis, HRTFs need to be individualized, matching the morphology of the listener. Typical means for this are to measure or calculate the HRTF of the listener, but these lengthy and costly methods are not feasible for general public applications. This paper presents a novel approach for HRTF individualization, separating the head and torso effect from that of the pinnae. The head/torso component is numerically modeled while the pinnae component is created using a multiple transducer array placed around each pinna. The philosophy of this method consists in trying to excite the correct localization cues provided by the diffraction of the reconstructed wave front on the listener's own pinnae. Simulations of HRTF reconstruction with various array sizes and preliminary auditory localization tests are presented.