A comparative study of LPC parameter representations and quantisation schemes for wideband speech coding

Switched Conditional PDF-Based Split VQ Using Gaussian Mixture Model

“…We also use another recently proposed measure of spectral distortion with interframe memory (SDM) [9]. A low average SD and rms SDM along with minimum number of high SD outliers are considered necessary for perceptually good spectrum quantization performance [2], [9], [13]. For better SD performance, the codebook is searched using the weighted square Euclidean distance (WSED) measure [2], [13], where the weights are the spectral sensitivity coefficients [4].…”

“…For each Voronoi region, the 16 dimensional LSF vector is split into five parts of (3,3,3,3,4) dimensional subvectors for both CSVQ and SVQ methods. The number of Voronoi regions and subvector dimensions are chosen according to [13]. The optimum parametric SCSVQ method is designed using variable bitrate scheme, resulting in fractional inter-cluster bit allocation [see (8)] and also fractional intra-cluster bit allocation [see (9)].…”

“…To recover the split loss, we exploit the conditional PDF and sequentially quantize each subvector , [15], [16]; i.e., , is quantized utilizing the past information . We have shown that the optimum CSVQ method provides the R/D performance as (from [16,Theorem 2]) (1) 1 These Voronoi regions are referred to as "switching regions" in [13] and [14]. In this letter, Voronoi region and switching region are used interchangeably.…”

“…This approach sacrifices the correlation between the subvectors and, thus, leads to a coding loss, referred to as "split loss" [10]. To recover the split loss, So and Paliwal [13], [14] have recently proposed switched split VQ (SSVQ) method, which provides better rate-distortion (R/D) performance than the traditional SVQ method. We also have proposed a parametric technique of conditional PDF-based split VQ (CSVQ) method to recover the split loss [15], [16].…”

“…An input vector to be quantized is first classified to a Voronoi region and then the region-specific SVQ is used for quantization. It was shown [13], [14] that an increase in the number of Voronoi regions leads to better recovery of the split loss, but at the cost of a steep increase in complexity. For a practically reasonable number of Voronoi regions in SSVQ, we note that the region specific SVQs are susceptible to split loss.…”