Arnaud Dessein

Real-Time Detection of Overlapping Sound Events with Non-Negative Matrix Factorization

In this paper, we investigate the problem of real-time detection of overlapping sound events by employing non-negative matrix factorization techniques. We consider a setup where audio streams arrive in real-time to the system and are decomposed onto a dictionary of event templates learned off-line prior to the decomposition. An important drawback of existing approaches in this context is the lack of controls on the decomposition. We propose and compare two provably convergent algorithms that address this issue, by controlling respectively the sparsity of the decomposition and the trade-off of the decomposition between the different frequency components. Sparsity regularization is considered in the framework of convex quadratic programming, while frequency compromise is introduced by employing the beta-divergence as a cost function. The two algorithms are evaluated on the multi-source detection tasks of polyphonic music transcription, drum transcription and environmental sound recognition. The obtained results show how the proposed approaches can improve detection in such applications, while maintaining low computational costs that are suitable for real-time.

Vocal Imitations and the Identification of Sound Events

It is commonly observed that a speaker vocally imitates a sound that she or he intends to communicate to an interlocutor. We report on an experiment that examined the assumption that vocal imitations can effectively communicate a referent sound and that they do so by conveying the features necessary for the identification of the referent sound event. Participants were required to sort a set of vocal imitations of everyday sounds. The resulting clusters corresponded in most of the cases to the categories of the referent sound events, indicating that the imitations enabled the listeners to recover what was imitated. Furthermore, a binary decision tree analysis showed that a few characteristic acoustic features predicted the clusters. These features also predicted the classification of the referent sounds but did not generalize to the categorization of other sounds. This showed that, for the speaker, vocally imitating a sound consists of conveying the acoustic features important for recognition, within the constraints of human vocal production. As such vocal imitations prove to be a phenomenon potentially useful to study sound identification.

An Information-Geometric Approach to Real-Time Audio Segmentation

We present a generic approach to real-time audio segmentation in the framework of information geometry for exponential families. The proposed system detects changes by monitoring the information rate of the signals as they arrive in time. We also address shortcomings of traditional cumulative sum approaches to change detection, which assume known parameters before change. This is done by considering exact generalized likelihood ratio test statistics, with a complete estimation of the unknown parameters in the respective hypotheses. We derive an efficient sequential scheme to compute these statistics through convex duality. We finally provide results for speech segmentation in speakers, and polyphonic music segmentation in note slices.

Online Change Detection in Exponential Families with Unknown Parameters

This paper studies online change detection in exponential families when both the parameters before and after change are unknown. We follow a standard statistical approach to sequential change detection with generalized likelihood ratio test statistics. We interpret these statistics within the framework of information geometry, hence providing a unified view of change detection for many common statistical models and corresponding distance functions. Using results from convex duality, we also derive an efficient scheme to compute the exact statistics sequentially, which allows their use in online settings where they are usually approximated for the sake of tractability. This is applied to real-world datasets of various natures, including onset detection in audio signals.

Perception of vocal imitations and identification of the imitated sounds.

We report two studies investigating how vocal imitations enable the recognition of the imitated sounds. First, we asked couples of participants to listen to series of everyday sounds. One of the participants (“the speaker”) had then to describe a selected sound to the other one (the “listener”), so that he could “guess” the selected sound. The results showed that, spontaneously, the speakers used, among other para‐linguistic cues, large numbers of vocal imitations. Moreover, they suggested that the identification performances were increased when vocal imitations were used, compared to only verbal descriptions. Second, we sampled 28 sounds across an experimental taxonomy of kitchen sounds and required laypersons to vocally imitate these sounds. Another group of participants was then required to categorize these vocal imitations, according to what they thought was imitated. A hierarchical cluster analysis showed that, overall, the categories of vocal imitations fitted well with the categories of imitated so...