This paper describes a new robust voiced / unvoiced classification algorithm, using an appropriate wavelet packet decomposition of speech signal. The classification is achieved by generating a correlation model of different subbands signals derived from a tree structured filter banks. The wavelet packet tree is constructed by cascading the basic two channel perferct reconstruction filters into the desired levels. To investigate the accuracy of the proposed technique, we conduct experiments using the TIMIT speech database. We add to these speech signals real word noise at various SNR. Experimental results show the accuracy of the proposed technique especially in low SNR's (<10dB).