Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
Advanced topics
and would apply these to each line in a set of LSPs. So for an original LSP set lsp, we can shift by degree b Barks to derive a set of shifted lines lsp2. We can shift line n with: lsp2(n)=bark2f(f2bark(lsp(n))+ b); However, it would of course not be necessary to shift every line in the set of LSPs, only the lines directly relating to particular formants would need to be shifted. The hard limit would still need to be applied to prevent LSP values approaching the angular frequency extremes of 0 or π and checks made to prevent unintentional resonances caused by moving two lines too close together. As already mentioned, anything other than small changes to a continuous speech spectrum is likely to lead to reduced quality. 7.8.2 Processing complexity In terms of processing complexity, LSP narrowing requires three operations for every line (nine operations for a typical three-formant frame). Shifting using Equation (7.14) requires four operations per line, or 40 operations to shift all lines in a tenth-order analysis frame. LSP shifting using Equation (7.15) requires around six operations per line but, when implemented, will usually necessitate a lookup-table. Similar formant processing effects can also be produced using adaptive filter tech- niques. Such a filter requires at least 2NP operations per N -sample Pth-order analysis frame. For tenth-order analysis operating on 240-sample frames, the LSP processes discussed here are between 40 and 400 times more efficient than an adaptive filter. Such figures are only valid where LSP data are available (which is the case for many CELP coders). If LSP data are not available, the overhead of transforming LPC coefficients to and from LSPs would be far greater than any possible efficiency gain. The methods of LSP adjustment described here have been successfully applied in the intelligibility enhancement of speech [26]. In particular, the use of LSP shifting for altering the balance of spectral power between formant peaks and valleys shows promise. Figure 7.9 illustrates the effect of increasing the separation between the three most closely-spaced line pairs by 20%, and was performed using the lspnarrow() Matlab function on page 190. The figure plots both the resultant spectrum (drawn with a solid line), and the original spectrum (drawn with a dashed line). Traditional bandwidth- altering adaptive filters, such as that of Schaub and Straub [27], perform a similar task, but at a higher computational cost. Further applications of LSP adjustment may include the quality enhancement of speech/audio and voice masking – described in Section 7.9. |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling