We propose a novel approach to language identification. Generally speaking, an ideal language identification system needs a large number of speech transcriptions at the phoneme level for training the phone model, involving a huge amount of work and cost. In this project, we use a rough segmentation instead of transcription to produce sub-words, and a front-end sub-words recognizer for individual languages to be identified. This is followed by clustering the sub-words and creating an HMM for each cluster. Preliminary results on language identification are provided to demonstrate simplicity and effectiveness of this approach.