This paper deals with the detection of the number of broadband sources in an acoustic environment from a small-size, embeddable microphone array. The coupling of information-theoretic statistical identification methods with MUSIC schemes, which has long been acknowledged in the array processing community, is first reviewed. From these considerations, a source number detector based either on the Akaike Information Criterion (AIC) or on the Minimum Description Length (MDL) criterion is derived within an original coherent broadband beamspace MUSIC method. This constitutes a theoretically sound solution, well-suited to robotics as it requires no subjective threshold setting and has a limited computational cost. Experimental results validate the approach. All the necessary theoretical background is provided, so that the paper is self-contained.