Abstract: A novel multiple description sinusoidal speech codec(MDSC) is proposed.This codec is based on sinusoidal and equivalent rectangular bands(ERB) noise model.The parameters of sinusoidal model are interleaved into two descriptions,and parameters of noise model are split vector quantized,and the indices are divided into two descriptions.The emulation results show that the model can provide relatively high transmission reliability as well as good coding efficiency.And the lost packet wouldnot affect the state recovery of this stateless MD codec.Therefore it is very suitable to unreliable and band limited channels such as the internet.