An extrapolation method for MDCT domain frame loss concealment

This paper studies the compressed domain extrapolation methods for error concealment in audio communication. Our work reveals that in the modified discrete cosine transform (MDCT) domain, traditional power extrapolation method is not proper for audio signals that have plenty of tonal components. By applying the sinusoid model analysis technique, a new receiver-based combined method is proposed to deal with the audio frame loss with no additional algorithm delay. A shaped-noise extrapolation is used to obtain the MDCT coefficients in noise-like spectral bins, while a sinusoid-model based extrapolation is performed to generate the MDCT coefficients in tone-dominant bins. Experimental results demonstrate that the proposed combined method provides better performance than the repetition method and the pure shaped-noise extrapolation method.