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
115 0 10 20 30 40 50 60 70 80 90 0 0.5 1 1.5 2 2.5 3 Angular frequency Normalised Count Figure 5.12 A histogram of the relative frequency difference between LSP line locations for consecutive analysis frames, plotted using 80 analysis bins. 1 and 10 exhibit less variation in their absolute value, but also in their frame-to-frame variation than do the centre lines. Experimental results suggest that dynamically choosing one or other of the two quan- tiser bases (frame-to-frame differential, or distance-from mean) to quantise and transmit can improve quantiser performance [9,16]. In practice this implies that the quantised parameters are calculated using both methods for each analysis frame, and the method which results in the closest match is then chosen. A single bit from the quantisation bit pool must then be reserved to flag which method is in use [9]. In a standard CELP coder employing scalar quantisation, around three or four bits are used to quantise each LSP [8,17], or around 30 or 40 bits for a tenth-order representation. A number of techniques have been published in order to allow representation with fewer bits at equivalent levels of quality. A survey of published results from [4,11] and other previously mentioned references can assess the spectral distortion associated with a number of different coding methods at between 3 and 3.6 bits per LSP. This is shown in Figure 5.13 which plots the spectral distortion due to quantising a tenth-order LSP representation using the following methods: • uniform scalar quantisation (UQ and LSF); • differential scalar quantisation (DQ and LSFD, also known as DF); • adaptive forward sequential quantisation (AQFW); • adaptive backward sequential quantisation (AQBW); • non-uniform scalar quantisation (NUQ); • dynamically interpolated quantisation (LSFI); • dynamically interpolated differential quantisation (LSFID). From the figure it can be seen that all methods improve at a similar rate as bit allo- cation increases, and that differential methods are better than non-differential methods. Furthermore non-uniform quantisation improves over uniform quantisation. It should |
Ma'lumotlar bazasi mualliflik huquqi bilan himoyalangan ©fayllar.org 2024
ma'muriyatiga murojaat qiling
ma'muriyatiga murojaat qiling