David Virette's research while affiliated with Orange Labs and other places

... Peak detection Peak detection is proposed in [28] to encode a single acoustic response with the goal of accelerating run-time convolution of input signals during playback of MPEG streams. Neither spatial interpolation nor compactness is a concern. ...

... In contrast to source-driven SCSS methods described in previous Section, model-based SCSS methods have been of high popularity due to their top separation performance. They rely on prior knowledge in the form of pre-trained dictionaries of the spectral amplitude information, using some statistical modeling techniques including factorial hidden Markov models [14,15], vector quantizers [16,17], graphical models [2], Gaussian mixture models [18], non-negative matrix factorization (NMF) [19,20,21,22], and deep models [23,24]. As our second proof-of-concept in this work, we select active-set Newton algorithm for overcomplete NMF recently proposed in [19], as it outperforms other conventional source separation techniques. ...

... • The phase-align principle, where the signals are temporally or phase aligned prior to the mixing process. This has been proposed, for instance, to improve parametric stereo coders [6,7,8]. A continuous and robust phase-alignment is not an easy task and any misalignment will immediately result in comb-lter artifacts. ...

... A well-known technique named multi-zone reproduction has been studied over the years [99][100][101][102][103][104]. The original idea of multi-zone reproduction is to present multiple individual sound fields at the target areas. ...

... Other works have considered log-regularization (i.e., penalizing the sum of the logarithms of the entries of the factor) which leads to more "aggressive" sparsity, e.g., [22]- [24]. Sparse regularization with information measures is also considered in [25], [26]. Another approach to induce sparsity consists in applying hard constraints to the factors (rather than mere penalization), using ℓ 0 constraints [27], [28] or using the sparseness measure introduced in [13]. ...

... The recommendation annex B (G.729B) [6] is designed with an idea of reducing average bit rate by using silence compression techniques for DSVD and other bit rate sensitive applications. The ITU-T G.729 based Embedded Variable bit rate (G.729EV) [8] provides a scalable coding scheme with narrowband to wideband audio quality in the range of 8-32 Kbps. There has been a significant work carried out about various schemes to reduce the bit rate and improving the speech quality and not much work has been reported regarding the techniques used for secured speech data transmission. ...

... The aliasing cancelation process of the proposed algorithm is conceptually similar to that of the block switching compensation scheme proposed for low delay advanced audio coding (AAC-LD [6,7]). In the literature, the scheme introduced time domain weightings applicable as a post processing in the decoder in order to remove a look-ahead delay inevitable for a window transition from the long window to the short window. ...

... As a result this hybrid configuration performs well for both speech and music signals. For more detailed results see [26]. Another hybrid example is the combination of sinusoidal and transform coding. ...

... In semi-supervised approach for real-time speech and background noise separation researcher used NMF, in which fixed speech bases are learnt from training data and noise components are calculated in real time using current data. Experiments with spontaneous conversational speech and real-world nonstationary noise reveal that proposed system outperforms a supervised NMF algorithm that learns noise components from the same noise environment as the test sample [32]. ...

... Primary ambient extraction techniques can be used for the task on decomposing a sound scene into separate objects. Principal component analysis is a popular technique for this task, whose performance depends on the correlation between the different components [134]. The decomposition task can be also performed by utilising blind source separation (BSS) separation techniques, assuming that the objects in the multichannel recording are mutually uncorrelated or statistically uncorrelated [135]. ...