Estimated lost LSFs: A new method based on intraframe CVQ compensation & interframe interpolation

To effectively estimate the linear spectrum frequency (LSF) by intra- and inter- frame dependence of LSFs, an intraframe conditional vector quantization (CVQ) combined with interframe interpolation (CVQ-I) is proposed to estimate the lost LSFs. We firstly partition the dependent 10-dimension LSFs in a frame into two bundles: the first 4-dimension vectors called low part LSFs (LPL, hereafter LPL also means 1 LSF in LPL) and the left 6-dimension vectors called high part LSFs (HPL). Then we enlarge the CVQ codebook so that each HPL maps a sub-codebook related to K LPLs. If a LPL is lost, K codevectors mapped to the same known HPL in CVQ are used as K estimations of the lost one, besides the (K+1)th estimation derived from linear interpolation between LPLs of the previous and the next available frames. The optimal LPL among these (K+1) estimations is finally obtained under the MMSE criterion. We also design an interleaving scheme CVQ-I which reserves LPLs while interleaves HPLs with those of the previous frame to recover a lost frame from the previous and the next available frames. Experiments show that PESQ of the new method is higher than interpolation by 0.017 at an average loss rate 10%. Its 0.005 promotion in MOS and a 0.0053dB decrease in spectrum distortion also demonstrate that adding several bits in codestream is a worthy pay to carry optimal indexes.