Applied Speech and Audio Processing: With matlab examples
Download 2.66 Mb. Pdf ko'rish
|
Applied Speech and Audio Processing With MATLAB Examples ( PDFDrive )
Speech communications
the static codebook storage is no longer required. For memory-constrained embedded systems, this may well be a significant consideration. 5.4.3 Split codebook schemes Unlike ACELP (above, Section 5.4.2) which reduces the computational complexity of the LPC filtering process, split codebook schemes attempt to reduce the ‘search space’. By default the search space is the entire codebook since each element in the codebook is tested in turn, and there may be very many such elements. If there were some way to reduce the number of candidate excitation vectors tested, then complexity can be reduced. Two main methods exist. The first one is to order the codebook in such a way that the codewords are arranged by known approximate characteristic. Through one of several different predictive processes, it is possible to determine the rough spectral or time- domain characteristic of the vector required for a particular speech analysis frame. Only codewords fitting the required description need be tested as candidates. Occasionally such a system may make a suboptimal choice, but given a sufficiently good predictive algorithm, such events may be rare. The second method is to use two or more different codebooks with orthogonal prop- erties. Each orthogonal property is found separately, and the two codebooks can be searched independently. As an example of a dual-codebook system, codebook 1 – per- haps carrying a selection of spectral distributions – is searched with codebook 2 set to some arbitrary value. Once the best match from codebook 1 is found, this is maintained whilst codebook 2 – perhaps carrying temporal envelope information – is searched. Ideally the two are orthogonal (meaning that whichever codeword from one codebook is used, the best match from the other codebook will not change for a particular frame of analysed speech). At one extreme, a split system could be a mixed excitation linear prediction (MELP) coder – although this name is typically reserved for coders with split excitation that do not utilise a codebook (and hence do not perform analysis-by-synthesis) [21]. VSELP (vector sum excited linear prediction [22]) is a common split system with ordered codebooks that are searched independently, with the best codeword from each added together to provide the excitation vector. In this case, each codebook output is subject to an individual scaling factor. Less important for this discussion, but a fact which should be mentioned for consistency, is that in the VSELP coder one of those codebooks is used to replace the LTP pitch filter. This codebook is pitch-adaptive, and is obviously searched, and optimised, before the other two codebooks of candidate vectors are searched. 5.4.4 Forward–backward CELP When using a speech coder such as CELP in a real application, it will naturally take some time for a particular sound of speech to travel from the input microphone, through the system, and arrive at the output loudspeaker. Apart from the transmission of speech through wiring (or fibre optics, or wireless), processing takes some finite time. |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling