Recent progress in corpus-based concatenative text-to-speech synthesis has generated some interest in systems that are capable of synthesizing text from more than one language. In this paper we describe the language identification component of such a mixed-lingual text-to-speech system. Relying only on the input text, we employ two different methods, namely a transformation based learning approach and a stochastic n-gram approach, and we describe the combination of both methods. While the transformation-based learning approach already produces average error rates of less than 2 percent and outperforms the n-gram classification scheme, the combination of both methods results in a further error reduction of up to 50 percent.