Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
5.2. Parameterisation
101 Figure 5.9 The role of LPC analysis and synthesis filters in many speech coding algorithms to convey parameterised vocal tract information from encoder to decoder. 5.2.1.2 LPC stability issues Typically the LPC parameters a ( ) are represented in Matlab as floating point values. If these were used in a speech coder to represent speech parameters then they would need to be saved as smaller fixed-point values for transmission between encoder and decoder (since floating point values occupy too many bits). Unfortunately this quantisation process can occasionally produce a set of parameters that results in an unstable synthesis filter in the decoder. This means that the LPC filter output magnitude rises sharply then ‘blips’ towards infinity before recovering. When listening to speech this effect would be recognisable as a loud pop or squeak. The same effect can sometimes be heard on live news broadcasts from remote locations by satellite phone – pops and blips in the audio disrupting a reporter’s voice are the result of errors in the parameterised audio bitstream ‘kicking’ the LPC synthesis filter into momentary instability. Of course, in the presence of bit errors in the parameters such things are expected, however even in the absence of errors, directly quantising LPC parameters often results in instability. We will look briefly at the effect of quantising LPC parameters directly in Section 5.2.5, but note that LPC parameters are never quantised directly – in practice they are always converted into an alternative form. One such form is the reflection coefficients described in Section 5.2.2. Log Area Ratios (LAR) are the method of choice for the GSM speech coder, comprising a logarithmic transformation of the reflection coefficients, and other more esoteric mathematical transformations exist. None of these can compete with the quantisation characteristics of Line Spectral Pairs (LSPs) which are worthy of their own discussion in Section 5.2.4. |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling