We present a new approach to the source separation problem in the case of multiple speech signals. The method is based on the use of automatic lipreading, the objective is to extract an acoustic speech signal from other acoustic signals by exploiting its coherence with the speaker′s lip movements. We consider the case of an additive stationary mixture of decorrelated sources, with no further assumptions on independence or non-Gaussian character. Firstly, we present a theoretical framework showing that it is indeed possible to separate a source when some of its spectral characteristics are provided to the system. Then we address the case of audio-visual sources. We show how, if a statistical model of the joint probability of visual and spectral audio input is learnt to quantify the audio-visual coherence, separation can be achieved by maximizing this probability. Finally, we present a number of separation results on a corpus of vowel-plosive-vowel sequences uttered by a single speaker, embedded in a mixture of other voices. We show that separation can be quite good for mixtures of 2, 3, and 5 sources. These results, while very preliminary, are encouraging, and are discussed in respect to their potential complementarity with traditional pure audio separation or enhancement techniques.

Looking at the speakerÕs face is useful to hear better a speech signal and extract it from competing sources before identification. This might result in elaborating new speech enhancement or extraction techniques exploiting the audiovisual coherence of speech stimuli. In this paper, a novel algorithm plugging audio-visual coherence estimated by statistical tools on classical blind source separation algorithms is presented, and its assessment is described. We show, in the case of additive mixtures, that this algorithm performs better than classical blind tools both when there are as many sensors as sources, and when there are less sensors than sources. Audio-visual coherence enables a focus on the speech source to extract. It may also be used at the output of a classical source separation algorithm, to select the ''best'' sensor with reference to a target source.

No abstract