R. Laroia, N. Phamdo and N. Farvardin, "Robust and efficient quantization of speech LSP parameters using structured vector quantizers," in Proc. ICASSP, pp. 641-644, 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the LSF parameters is a very high inter frame correlation. This high correlation indicates that the LSF parameters of a given frame can be predicted from the LSF parameters of the previous frames. Inter frame LSF coders use this property to attain performance improvementover the intra frame coders [25, 27]. Since inter frame encoders use the information of the previous frames to encode an LSF parameter of a certain frame, they all suffer from the propagation of errors for communication over noisy channels. Ohmuro et al. considered a Moving Average (MA) prediction scheme for differential ....

....quantization of LSF parameters in the formant regions than those in the non formant regions. Paliwal and Atal in [21] suggested assigning a variable weight w i to the ith LSF which is proportional to the value of LPC power spectrum at this frequency. Another simpler weight function was proposed in [25] whichtakes advantage of the fact that formant frequencies are located at the position of two or three closely located LSF parameters. Equation (3.19) is the definition of the metric used in this work. Due to recommendation of NORTEL ,wechoose constantweight c = 1# 1#: #1] and employ a ....

R. Laroia, N. Phamdo, and N. Farvardin, "Robust and efficientquantization of speech LSP parameters using structured vector quantizers," in Proc. Int. Conf. Acoust., Speech, Signal Processing, (Toronto), pp. 641-644, 1991.

....Kp 1 vith the Levinson reoursion: Ap 1 (z) Ap(Z) Kp 1 z (p 1) Ap(Z I ) 9) The odd and even LSP polynomials P(z) and Q(z) are formed by settin 8 the (p l) th reflection coefficient to 1 or 1 respectively in eq. 9) and can be written as: z 1 z 2) P(z) 1 Z I)H (1 2Y2i 1 i= 10) Q(z) l z I)II (1 22 i z l z 2) il These expressions are given here for even order LPC models. We only consider this case here for ease of notation, without loss of generalits, The LSP coefficients Ts are related to the LSP frequencies to i as ti= cos(C0i) By expanding the polynomials ....

....(a) b) c) the approximation of the Spectral Distortion of eq, 7) tbr ISP, 1SP and LAR, d) a WEDM for LAR. using as weighting factors the single parameter spectral sensitivities, i.e. only the diagonal elements of CL t 1, e) 1) g) WEDM for LSP using the heuristic weights in [21, 101 and [111. The standard deviations of the measures were: a) 0.02 )0.03 (c) 0.04 (d) 0.26 (c) 0.15 (00.08 (g007 respectively. The WEDM for LAR has vorst performance, because the off:diagonal elements of CLAR 1 are neglected. 20 30 Figure 1. Power Spectrum and LSP tYequeneies of an LPC ....

R. Laroia, N. Phamdo and N. Farvardin, "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantization", Proc. ICASSP, pp. 641-644, 1991.

....relate closely to formant frequencies (Crosmer 1985) but in contrast to formant frequencies they can be estimated quite reliably. They have been used for a number of applications successfully in the literature (Hansen and Clements 1991, Arslan et al. 1995, Arslan and Talkin 1997, Crosmer 1985, Laroia et al. 1991, Itakura 1975, Pellom and Hansen 1997) They have good interpolation properties and they are stable (Paliwal 1995) In addition, they have a fixed dynamic range which makes them attractive for real time DSP implementation. LSFs can be estimated by modifying the LPC polynomial, A(z) in two ways: ....

....centroid, S i , in the source codebook and the distance, d i , corresponding to each codeword is calculated. The distance calculation is based on a perceptual criterion where closely spaced line spectral frequencies which are likely to correspond to formant locations are assigned higher weights (Laroia et al. 1991), h k = 1 argmin(jw k Gamma w k Gamma1 j; jw k Gamma w k 1 j) k = 1; P d i = P X k=1 h k jw k Gamma S ik j i = 1; L (7) Based on the distances from each codebook entry, an expression for the normalized codebook weights can be obtained as (Arslan et al. 1995) v i = ....

R. Laroia, N. Phamdo, and N. Farvardin (1991). "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers". In Proc. IEEE ICASSP, pp. 641--644.

....into 4 sub vectors of dimension 4 each and each sub vector is vector quantized with 6 bits. LSFs are then represented by 26 bits, i.e. 1.675 bit LSF frame. In the split VQ codevector searches, the norm of the error vector, which we attempt to minimize, takes into account the distance between LSFs [9], the signal to mask ratio spectrum, and the decrease in frequency resolution for higher frequencies. The original speech signal and the STP filtered excitation signal are weighted with a perceptual filter of the type ) 2 1 g g = z A z A z W where A(z) is the STP filter. The ....

Laroia, R.; Phamdo, N.; Farvardin, N., "Robust and efficient quantization of speech LSP parameters using structured vector quantizers". ICASSP 91, 5, 641-644

....calculated. The distance calculation is based on a perceptual criterion where closely spaced line spectral frequencies which are likely to correspond to formant locations are assigned higher weights. The weights of the line spectral frequencies are calculated based on the formulation proposed in [6], hk = 1 argmin(jwk Gamma wk Gamma1 j; jwk Gamma wk 1 j) k = 1; P d i = P X k=1 hk jwk Gamma S ik j i = 1; L (2) where L is the codebook size. In addition to above weighting, for voiced segments, lower order LSFs, and for unvoiced segments, higher order LSFs are weighted ....

R. Laroia, N. Phamdo, and N. Farvardin. "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers". In Proc. IEEE ICASSP, pages 641--644, 1991.

....vector quantizer. However, due to complexity and storage constraints, only a relatively coarse, ten bit quantizer was used, and extending this work to higher rate quantizers is difficult. Vector quantizers which minimize the MSE in the LSP parameters have been previously proposed (e.g. 53] In [54], the fact that the sensitivity of an LSP frequency increases as the distance to the neighboring LSP frequencies decreases was used in a VQ system. In that work, the VQ was trained by minimizing an WMSE measure with weightings proportional to the reciprocal of the distance to the neighboring LSP ....

....at a slower rate per bit than for the other 2 subvectors, since the third subvector has a larger dimension size. In fact, over half of the LSD at high rate comes from the quantization of the highest 112 Error Measure Low 3 LSPs Middle 3 LSPs High 4 LSPs 10 LSPs MSE 0.249 0.371 0.788 1. 407 [54] 0.217 0.319 0.725 1.260 [56]a 0.223 0.333 1.074 1.629 [56]b 0.734 1.288 Optimal WMSE 0.198 0.298 0.699 1.194 Table 6.1: Comparison of LSD in dB 2 of 30 bit Vector Quantizers 4 LSPs. Also, the sum of the three approximations does offer a good approximation to the overall VQ performance at high ....

[Article contains additional citation context not shown here]

R. Laroia, N. Phamdo, and N. Farvardin, "Robust and Efficient Quantization of Speech LSP Parameters using Structured Vector Quantizers," in Proceedings of ICASSP, 1991.

....robust VQ was applied for the quantization of line spectrum frequencies (LSF) LSF are parameters that describe a linear predictor, here of order 10, in the frequency domain. The LSF can be coded very efficiently using vector quantization, especially if a weighted distortion measure is used [9]. At first the LSF were quantized by 10 dimensional full search vector quantization (FSVQ) with 1024, 2048 or 4096 codebook vectors. Then split codebook vector quantization (SCVQ) consisting of vector quantizers with dimension 2,4,4 was investigated [10] As a measure for the quantization accuracy ....

R. Laroia, N. Phamdo, N. Farvardin. "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers". Proc IEEE Int. Conf. on Acoust., Speech, Signal Processing. Toronto 1991.

....on the human perception of acoustic signals than similar formants that have wider bandwidths. The distance calculation takes this perceptual weighting into account, and closely spaced line spectral frequencies which are likely to correspond to sharp formant locations are assigned higher weights (Laroia et al. 1991): h k = 1 argmin(jw k Gamma w k Gamma1 j; jw k Gamma w k 1 j) k = 1; P d i = P X k=1 h k jw k Gamma L ik j i = 1; 5N (6) where 5N is the reduced codebook size based on context. However, if the phonetic information is not available 5N can be replaced with the whole codebook ....

R. Laroia, N. Phamdo, and N. Farvardin (1991). "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers". In Proc. IEEE ICASSP, pp.

....both in the amount of computation and in the size of memory required. We use a VQ codebook of LSF s of size 256. We calculate the distance of the noisy frame LSF s from each of the codebook entries. The calculation of the distance is based on a perceptual weighting called the inverse harmonic mean [6], and is given by: dk = P X i=1 w i (LSFni Gamma LSFki) k = 1; 256 w i = 1 LSFni Gamma LSFnc where dk is the distance corresponding to the k th codeword, P is the number of LSF s, LSFn refers to the noisy frame LSF s, LSFk refers to LSF s from the k th codeword, w i is the ....

R. Laroia, N. Phamdo, and N. Farvardin, "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers", in Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, pp. 641-644, 1991.

....context phone is represented with 5 uniformly spaced audio visual vectors. The incoming LSF vector w is compared with each LSF vector, L i , in the codebook and the distance, d i , corresponding to each codeword is calculated. The distance calculation assigns higher weights to closely spaced LSFs (Laroia et al. 1991): hk = 1 argmin(jwk Gamma wk Gamma1 j; jwk Gamma wk 1 j) k = 1; P d i = P X k=1 hk jwk Gamma L ik j i = 1; 5N (5) where P is the LPC analysis order, 5N is the reduced codebook size based on context. However, if the phonetic information is not available 5N can be replaced ....

R. Laroia, N. Phamdo, and N. Farvardin (1991). "Robust and Efficient Quantization of Speech LSP Parameters Using Structured Vector Quantizers". In Proc. IEEE ICASSP, pp. 641--644.

No context found.

R. Laroia, N. Phamdo and N. Farvardin, "Robust and efficient quantization of speech LSP parameters using structured vector quantizers," in Proc. ICASSP, pp. 641-644, 1991.

No context found.

R. Laroia, N. Phamdo, and N. Farvardin, "Robust and efficient quantization of speech LSP parameters using structured vector quantisers," in Proc. ICASSP'91, Toronto, Ont., Canada, May 1992, pp. 641--644.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC