In this paper, the problem of the design of a simple and efficient music-speech discriminator for large audio data sets in which advanced music playing techniques are taught and voice and music are intrinsically interleaved is addressed. In the process, a number of features used in speech-music discrimination are defined and evaluated over the available data set. Specifically, the data set contains pieces of classical music played with different and unspecified instruments (or even lyrics) and the voice of a teacher (a top music performer) or even the overlapped voice of the translator and other persons. After an initial test of the performance of the features implemented, a selection process is started, which takes into account the type of classifier selected beforehand, to achieve good discrimination performance and computational efficiency, as shown in the experiments. The discrimination application has been defined and tested on a large data set supplied by Fundacion Albeniz, containing a large variety of classical music pieces played with different instrument, which include comments and speeches of famous performers.