TW507194B - Variable-rate residual-transform vocoders using auditory perception approximation - Google Patents

Abstract

In this invention, a simple variable rate vocoder for compressing the speech with low computational complexity is proposed by exploiting an adaptive pitch predictor, discrete cosine transform (DCT) on its residual signal, and perceptual representation of the DCT coefficients. To improve the speech quality, we further suggest pre-emphasized and de-emphasized filters respectively figured in the encoder and the decoder. The proposed speech coder, which achieves a reasonable quality but with much lower computational complexity than the traditional coded excited linear prediction (CELP) coders, can simultaneously offer multiple bit rates with various speech quality in the same coding structure. The real-time implementation comprising both encoder and decoder in a single low cost DSP chip confirms that the proposed vocoder produces an acceptable speech quality. With simple and variable-rate advantages, the proposed vocoder is suitable for providing digital speech communication over PSTN, internet, and wireless networks to achieve quality-of-service applications.

Description

507194 Case number 89110061

V. Description of the invention (1) Background of the invention In the era of multi-media communication, following the trend of __ roads and mobile radio = if the frequency is too high, we can move down. The two βσ Η αα quality can eliminate congestion at a low rate. To prevent the interruption of communication. In addition, the International Motion Picture Experts Group 0 ™) group "the importance of ordering codec systems can also be valued and developed. Including the frequency agreement, the low-rate 1-reduction speech codec system currently used, is 6 The systems are very large. Although it has a bit rate of 6.3Kbps, but the codec compression bit rate is not changed, the core structure is not exactly the same, and it cannot be used; the standard bit The rate is adjustable, but the complexity of the entire architecture and the 1Ti G · 723 · 1 system. The actual application of multimedia communication is the same as that seen on a DSP or CPU chip. In the multimedia communication system of the conventional technology, transmission Bandwidth will change greatly with time and environment. Therefore, it is necessary to check the development of the Internet and Internet and reduce the use of bandwidth. From 1975, many edits Maji can produce high-quality speech at a low bit rate, most of them == I pay off page 8.2002.05.08. 〇〇8 507194 'SL · 1

Case No. 89110061 V. Explanation of the invention (2) 丨 Explanation (2) 〇〇4 ^ ί .Α, Therefore | Lion ^^ || Fiber optic I certificate f | J? Quantitative quantity, and use the closed loop search county to get the excitation signal This H, using closed-loop search cleverly requires a large amount of calculation, even if it is a program of coding and decoding. As for the application of variable rate in today's literature, the ::: mb encoding program needs to use codebooks of various sizes, so: In 1993, Y · Shoham Proc · ICASSP-93 II-167-T69 ¥ frequency, the field of conversion to shrink the speech, the conversion coding method belongs to the voice quality of i 6f. Among them, the discrete cosine transform (D⑺ is not only a Σίΐϋ type, encoding and decoding technology. That is, by changing it to edit ^ ^ to easily delete some less important dct system = Κ2ίί sound quality, but at = ¾ 率 Ϊ; Αί4ίί: ί2ί To build a signal based on the characteristics of the sound, a strict invention target system will be generated. A new variable rate speech coding system structure is used. Ί5ΐΐ = 3 does not need to change the system when the rate is changed. 9Page 9 2002.05.08.009 507194 Case No. 89110ΠB1 V. Description of the invention (3) code. In order to further improve the speech. For the time being, to emphasize the two wave filters, the two f-wave system codecs use pre-emphasis instructions. Liu Yu action without wheels, detailed description of the invention iSSii special code Yi forget garden code ί5ίί, the invention uses a DSP chip and a personal computer; now it is clear that "an auditory characteristic is used for residual value conversion variable rate speech codec A device that uses pre-emphasis filtering of auditory characteristics to convert residual values at the base period to predict the "ratio" speech coding and its corresponding decoding and de-emphasis filtering device. The device includes: c pre-emphasis filter tail · e (Z) · ·will The original input speech signal 5V (/ 7) is processed into a pre-emphasized speech signal S〇7); the filter coefficient is obtained by using the previous / decoded sound frame speech signal and obtained by the corpse-level linear predictor (LPC) Coefficients, {α /, ΙΒΙ, page 10 20〇2.05.08.010 507194 _ case number 89110061

Jt. Amendment V. Invention Description (4) 1, 2, ..., household}, where H pre (Z): 1 * Σ α / 6 / ζ- /-Qf device: This title uses an adaptive codebook (CGdebook) Calculate the period parameter view by J soil square difference 1, and the gain parameter is the sum of the pre-strong speech signal 5㈤ minus the above-mentioned period reference parameter to obtain an estimated signal. (4, to get a residual value signal x \ n); —? Converter (DCT) ·· Receive the residual value signal of the above-mentioned subtractor Can, take «points as an input sound box, overlap # in time domain, create # points Convert 1 frame, and then convert it into #point DCT coefficients and 々]; input frame time domain (Kin must = 刖 take overlapping #points to create a #point conversion frame, a frame window function (Frame Window Function) «{ y;] = i, M < n <; ήη]-(2/7 + 1) / 2i (0 < / 7 <#-i; ^ π] = [2 (^ / 7) -1] / 2i ( ## < /? <#-/; —Coefficient Picker: Receives the aforementioned allowable error of the heterogeneous DCT coefficient system based on the characteristics of person f and the frequency variation of people, and the DCT coefficient f, the coefficient of the point of view in the frequency domain Divided into five different parsing regions, the f-feeling characteristics are selected from the absolute values of f systems and numbers to form the DCT coefficient vector.

Ra = te) k ^: T series: BaSeRate coding and high rate (High dagger: the selected base rate DCT coefficient vector is calculated as a weighted vector, i HHUan ^ Zati〇n) Obtain the best codeword index (°° ㈤k ㈤ 参, 巧 wonderfully good, the word quantization gain (Q_tized codeb〇k Gain) P3 'The above-selected high-rate DCT coefficients are selected based on the quantized coding base.

Page 11 2002.05.08.011 X507194 Year Flat 59

V. Description of the invention (5) The number is used as a reference for scalar quantization (Scalar Quantization); where the DU coefficient picker must be divided into five segments A, β, c, D, and E in the frequency domain. The area is 250 Hz, the C area is 500 Hz, and the D area ^ 疋 ΙΚΗζ 'E area is 2KHz and the like. In the basic rate coding, the maximum absolute DCT value coefficients in each of the four regions, A, B, C, and D, are coded with an increasing number vector; at a low rate ((L0wRate), the coding is increased. Each additional point of DCT coefficients in the two areas of a and β (that is, on the architecture of this edition, another area of a is added to select the individual additional encoding point system based on the auditory occlusion effect as a reference and is normalized with the largest DCT value in the area. When using 3 bits to purify in the Middle Rate, add an additional DCT coefficient in each of the two regions of coding d (that is, on the architecture of this edition, add two additional regions c and D to The auditory occlusion effect is used as a reference to select individual additional coding coefficients, and after the maximum DCT value in the area is normalized, it is quantized in 3 bits. At high rate, the coding is increased based on the medium rate. 4 points of DCT coefficients in the area (that is, based on the structure of this editor, and additionally added the occlusion effect based on the auditory hearing, and 4 bitcoin codes of 4 DCT coefficients are extracted by 3-bit quenching), as shown in Table (1), where The position index is to select the one with the largest sum of the DCT amount of the corresponding position, and its size is 4 bits. Which codebook index code, the codeword big ": The quantity of pure chance to gain bit depth using the encoded quantization and the like:

ssasassasaas position index position 1 000 32,3M4,50 area E 001 34,40,46, 52 010 36,42, 48,54 011 39,45,51,57 L100 141,47,53,59 101 43,49, 55,61 110 45,51,56,62 111 48,55,58,60 I 1 · ΗΙ Page 12 2002.05,03.012 507194 Case No. 89110061 Take 6 years! Month ... Miscellaneous V. Description of the Invention (6)- One-parameter multiplexer ·· Integrates the above-mentioned period parameters # and, the period gain parameters are 々, DCT codeword index, 1 € 1 quantization gain &, and quantization of DCT coefficients added due to high rate, etc. Data to form a compressed data stream to complete multi-rate speech compression; one: a parametric demultiplexer: separates the above compressed data stream into period parameters; # and,, 'period gain, number, and /, DCT codeword indexing , # DCT positive and negative values, DCT quantization gain &, and quantization of DCT coefficients added due to high rate, start speech decompression; a discrete cosine transform composer: the above parameter demultiplexer solves DCT codeword index is sparse, // DCT positive and negative values, and DCT quantization gain &, first perform #point build, and then send additional DCT The inverse quantization of the number, which reconstructs the DCT coefficients of the water point with a zero value reconstruction; an inverse ϊ1 cosine transform (IDCT) device, which replaces the reconstructed water point DCT coefficients with the signal of the residual value of the reconstructed sound box, and The front noise is added to the previous reconstruction sound box to reconstruct (continued point residual value signal; K mouse points are phased, the generator is estimated, the above parameters are solved by the multiplexer and the period gain parameters and / Produces exactly the same 编码 1; 'yes': the adaptive codeword of /. It is related to c •: its code = benefit P !, P2'U produces an estimated speech vector & its relationship The formula is as follows: ~ = Piq + p2Cy •, where A / = ί · ± 1, and the best base period is ^ (5C /) 2. / C / C /: Conversely / k =: / 氺 +1 Yuan Jiaji Zhou Parameter y * = / *-1 as (SC.f ^ (sc.) 2 9-, 9-, ~ 9 + Ι9 + ι 9-ι9-ι 9 +, 9 + 1 must be summed with least square error It is found that the best one is HBHil. Page 13 2002 · 05.08.013 507194 In the revised year and month, "Yue, Invention Description (7) The quantification of quantification is equal to the probability of a scalar quantification. For the long quantification, the first Transform, and then quantize the 〆 scalar to β. One adder: reconstruct the above Frame residual value signal; and the estimated signal? (Π) are added to reconstruct the pre-strengthened speech signal +); the filter is emphasized ^ 5e (Z) ·· The above-mentioned reconstructed pre-strengthened speech signal f (/ 〇 is processed into Finally, the speech signal is reconstructed to complete the decompression of the speech. The following describes the related technical structure of the present invention as follows (a). The codec block diagram (a) and (b) are the auditory characteristics proposed by the present invention. Basic structure of speech encoder and decoder for residual value transition rate. The coding algorithm proposed in the present invention is based on a frame with less DCT values as the coding core. In order to reduce the noise caused by the Biocky Effect during DCT conversion, the present invention has # The speech samples of the points overlap, as shown in Figure (3). In order to reduce the noise generated during the DCT conversion, the present invention uses pre-emphasis 100 and de-emphasis filter and wave filter 180 for processing at the codec side, as shown in (a) and (b). The original speech is processed by a pre-emphasis filter and then the base-period estimated signal is subtracted to obtain a residual value signal, and then the residual value signal is converted into DCT coefficients by discrete cosine transform (DCT) 110. g uses the characteristics of human psychoacoustics ears, and selects a larger audible (Audible) DCT coefficient for a small number of DCT values of the sound frame to quantize the encoding. At the solution end, the inverse vector quantization and inverse DCT conversion (brittleness) 130 are obtained. The residual value is added by the estimated signal 260, 180, and paid to # points of synthetic speech. The above details are described below: Round Page 14 2002 · 05 · 08 · 014 507194 Catch 5. · 13. Name) Multi-year i-89110061 V. Description of the invention (8) (2) Second-order adaptive estimator standard, the adaptive estimator is used Least square error standard 140, 128 codewords are stored in the format of 二 2 / recommendation codebook Fengli Stealth®. These characters are overtone signals, and the codebook is from 0 · to 127 ni base cycle delay '㈣ sampled from 20 to 147, that is, the base cycle delay, index 1 corresponds to the base cycle delay of 21, and I = the base cycle delay is approximately the same, this assumption is reasonable. For non-integer base cycle delays, the present invention ... ^ i = PlC / + P2Cy, for j = i ± l ⑴ Mainly continuous ㈣ characters whose codeword gain can be seen in the estimated signal vector center by two The physical meaning of a consecutive codeword is that the two codewords are interpolated. Therefore, the day t 疋 belongs to the main base cycle delay, and the index ′ is the second soil delay. In order to reduce the calculation amount, the search mainly refers to the first-order estimation standard, as follows: ^ J? AT ^ 1 (2)% green ban, 一 / C / C / ίϊίί emphasizes the sound frame signal, calculates the marriage value with u 1¾ in the codebook, and calculates = if ㊇.- 丨) yp / * + 1 if cr- icr- \ Ci \ icr + iς · + Ις · + Ι (3) ΙΒΙ Page 15 2002 · 05.08.015 507194 V. Description of the invention (9) ίίί! = ί5ί: ί? ί Sum *, the present invention | | E || 2 = IIS-plCr-P2Cyll2 —y2 is also 2 邛%, gas · + 2p, w ⑷ I is partially differentiating A and A individually, and let the equation be zero to get y5), ⑹2 ^ (β2 || €, || 2 ^ 0 / + βι € ι, €,) =: 〇, 5βι mi! 5β2 ^ C_ = 2n3, IIC..II2U, r- 广, Λ (5) σΡ2 JJ · I j, ”最佳 The best following the second base cycle gains 1 · and A · which are represented as follows ^ ^ (SC.XIIC.-I ^^ CX ^ CSC,) 1 one (7) = (8 ^) (|| < ^ || 2)-((^ ·)% ·) 2 (8) = ¾Wen, so the coding is as follows: -order estimator, but the second i-base is 4¾¾¾, and then the machine face quantity is quantified. , So the miscellaneous _ value vector becomes the following formula after the second-order estimation process (和 and 〆 are quantified for and / respectively) ... X ~ s ~ ^ m -ppjCy. (9) (3) Pre-emphasis and de-emphasis filter sound quality 'The voice signal at the encoding end sound box is pre-emphasized After subtracting from the adaptive prediction signal, the point residual value can be seen and then converted into the point DCT coefficient, so the inverse DCT conversion on the decoding side is significant. Page 16 2002.05.08.016 507194 Case No. 89110061 V. Description of the invention ( 10) Month a Corrected the clever frame yelling _ Chu Zhouyin she _ plus get # point frame excitation signal and then emphasize the filter 180 processing to obtain the transfer function of the following two formulas: 丨-Σ

Hde (z) aibi ^ μ-Σ a / W '/ = 1_ ρ ---- Σ (10) (11) * JLa ^ I' * ^ ^ Going strong_, wave ㈣ ') is a good thing: Frame, the audio signal is combined into a point signal and decoded by the meaning of 丨 (). It can be seen that the response of the new signal varies with every 2 2! Day signals, so it can be filtered out due to quantization or other factors. The weighting factor 5 of the wave filter can be selected as required. Since the parameters of the air compressor are obtained by decoding the sound of the voice frame, the parameters of the present invention to emphasize the filter need not be transmitted by the encoding end. (4) Quantification of auditory characteristics DCT coefficients vu and f improve the quality of synthesized speech. The present invention uses auditory characteristics to transform the ^ cosine coefficient, because DCT conversion has the best energy concentration, Γτγ 1 and eight different The amount is also less than the best KLT and there are already a lot of fast-you have to change the conversion formula of DCT and IDCT in the following 5! God. {Τ:

Jt = 0 did: ί 1 fox 1 / V2 forA: = 〇5 (12) (13) (14) Page 17 2002.05.08.017 507194

And especially [0] is the seventh sample of the residual signal, the DCT coefficient. It is the first one based on the statistical characteristics of auditory characteristics and DCT coefficients. The present invention uses encoding technology to #quantize these DCT coefficients. Because the present invention must be partially encoded, one part is the encoding frequency index, and the other is the relative quantization size. S is further explained as follows: DCT coefficient frequency index coding | Firstly, the energy dominated by the low-frequency components in the frequency analysis is sensitive to the higher frequency components, so the quantization and coding of the low-frequency DCT coefficients will use more bits, and the DCT coefficients of the south frequency components will be used. Fewer bits. In the field of energy conversion, the concept of low-frequency components and higher-frequency components is important. At the same time, the human ear has a hearing characteristic and the low-frequency range is more sensitive to frequency offsets. Therefore, the present invention has different frequency resolutions to encode DCT coefficients in different frequency ranges. Since each sound box has N points of residual signals, the DCT coefficients of the residual signals are also N points. According to the frequency index, the present invention divides the n points of DCT coefficients into 5 different parsing regions. The division method is shown in Figure (4). It is shown that the DC ^ coefficients in different regions will be encoded with different numbers of bits. According to the occlusion effect of human hearing, the energy component will occlude its adjacent low-energy components, so the present invention selects the largest energy in each region]) CT coefficients to encode, the frequency tolerance auditory tolerance is another An important psychological auditory feature, this Luminus-encoded DCT coefficient will also use this important feature. It is directly said that the human auditory system is sensitive to the shift of low-frequency components and higher frequency components. For the sampling frequency of 8KHz, the N-point DCT conversion region corresponds to the 4Khz frequency band. Because the DCT coefficients in the low frequency band are more sensitive and important to the auditory system, they are divided into smaller regions. The division is shown in Figure (4). A and The two areas of B are 250Hz, the C area is 500Hz, the D area is IKΗζ, and the E area is 2KHz. The present invention can selectively circle according to the size of the available bandwidth and the required voice quality_1 page 18 2002.05.08.018 507194

^^ 89110061 V. Description of the invention (12) Coding DCT coefficients in these five segments. For example, at 5.4 kbps, only the 4 maximum DCT values in the four regions of BC 'D are encoded. When quality requirements are raised, first, In the two areas A and B, according to the masking effect of human hearing, the original maximum energy in the area ^ is used as a reference, and the DCT coefficient of a relative position is additionally coded according to formula (15) to increase the quality of its synthesized speech. For example, 対 2] is The largest DCT coefficient in Henley in the luxury area is the additional DCT coefficient, $ 0], which is 6.4 k bit rate coding. If the required speech quality is still J, it is added in the C and D areas with the same coding method in the A and B areas. The DCT coefficient is 7.5k bit rate. Based on ;; = ps, r5 kbps, increase the 'N / 2 points out of the group in the E region based on the auditory shadowing effect to have 4 DCT coefficients. These 8 groups are selected to quote the' I ^ CT imitation coded ' ; As shown in Table 41: k ^ k-h2 k =: 0,1,4,5 k ^ k-2 k = 2,3,6,7 U > k ^ S I5 > k > l2 23 > k > \ 6 31 > / t > 24 k = ^: + 4? = Free + 8 k two) c one% (15) B, vector and adaptive quantization of DCT amplitude

iL5: Abf coding technique, the frequency cable selected according to the aforementioned principle. The ρ coefficient coefficient amplitude is encoded and transmitted to the decoder. • The white number of the remaining A is weighted to quantify it. It is an extremely large weighting factor. In the conversion field, it belongs to the larger π prime, so it is weighted larger. With the inverse gradual use of inverse. DCT transformation 130 reconstruction was 507194 Case No. 89110061 De 5, J3, ί3 J: _El said # Zheng V. Description of the invention (13) Plus the weight of the yT coefficient is gradually reduced, for convenience and Efficiency, the amplitude of the DCT coefficient is expressed as the magnitude of the absolute value and with positive and negative signs. In other words, the code of Cangbei 'No ί' lets X [k] be the range of k of the DCT frequency index k from 0 to N-1, For example, suppose that the encoded frequency index k is to be expressed as Kun 1 and Yang 4 to 5], in particular, the value of the dct coefficient f value: called = K 丨 '~ is the kth code of the DCT size copy, weighted by The expression of vector quantization and measurement distance is as follows: 4dk = Zwj (xj-p3mkj) 2, (16) g is the error standard weighting factor, ^ the gain of the DCT codeword amplitude, the characteristics of the weighting factor, the weighting factor indicating the lowest frequency, From the smallest '' ... one, this, *, take m plus thousands plus: distance, the present invention can get • matching codeword, especially as opposed to most of its type Max gain long-tailed bird is described as: ^ the following formula "kj (17) Ρ3-Σ; X;

Vj (18) i = CT4fii, ff f5 is 4. This is coded with 5 bits. For example, assuming that the rate is positive ^ = 2 2 _ page 20 2002.05.08.020 507194 case number 89110061 day I MU! Explanation (14) The amplitude of the added DCTX [0] is normalized and expressed as, since the brush is the largest DCT value in the region A, it will be less than 丨, so [[]] 丨 is easy to express by heart = 1 ^ Ka uses three bits to do qualitative quantification with equal probability. The reconstruction of Bu [〇ϊ is represented as 丨 Yebu • 丨 checkmark, the expression & is quantified, and its bit configuration is shown in Table 2. The 7.5Kbps coding technique is based on the 6.4Kbps codec, but the DCT coefficients added in the C and D regions have the same amplitude coding method as the a and B regions. The 9 · 4kbps encoding method is based on a 7 · 5kbps encoder. A β area is newly added, and 4 DCT coefficients are taken out to form a vector codebook with a dimension of 4, and then its codebook index is encoded with 4 bits. The word size gain is coded with a 3-bit parsing probability and scalar quantization.

(5) Real-time (real-time) realization and objective quality evaluation t; π Ί estimates the synthetic speech quality f and the calculation is complicated. The present invention integrates the base rate = device and the Fanma device at the same time in the same TMS320C30 to perform the conversion, and 4.3. The basic rate coding of the headline ㈣ complexity only requires ϋϋ 三 ί 5. 4Klys complexity evaluation, as for the evaluation of the quality of synthesized speech, and the ratio of signal to noise per frame, the rate of speech suppression and traditional 4._s fΐ 二 ίiif ί Early system calculations have a small amount of calculations. From Table 4, it can be seen that the four real _bit rates are better than H4.8H. Page 21 2002.05.09.021 507194 Case No. 89110061 Rev. V. Invention Description (15 ) The variable-rate speech coder only needs a small amount of additional calculation when changing the transmission bit rate, and does not need to change the architecture of the entire system, so it is very suitable to be implemented with a low-cost DSP hardware chip.

II ίΒΒ Page 22, 2002.05.09.022 507194 91.5, 13 Case No. 89110061 Brief description of the correction method Table 1: Table 1 DCT position encoding table for area E Table 2 Bit allocation table for the four-segment rate of the present invention Table 3 The calculation complexity analysis table of the variable rate speech coder according to the present invention is shown in Table 4. The quality comparison table of the variable rate speech coder according to the present invention is illustrated in the first diagram, the second diagram, the third diagram, and the fourth diagram. schematic diagram. Schematic diagram of the structure of a variable rate speech decoder. Schematic diagram of frame window function. 64-point DCT is divided into five different analysis frequency regions. Figure number description: 100 ... Pre-emphasis filter Hpre (z) ... Subtractor discrete cosine transform (DCT) DCT quantized inverse discrete cosine transform (IDCT) Base period estimation gain 扈, β adaptive codebook residual signal memory speech memory, analysis of sexual estimation coefficients… ^ emphasis filter Hde (Z) • • parametric multiplexer • • parametric solution multiplexing • • addition Multiplier ... Page 23, 2002, 05, 09, 023

Claims (1)

------ 89110061_year ", 1 # · 28: Take the eighteenth, apply for the scope of patent I— ________________:" L two & ^ audio auditory characteristics are used for the residual value Han Changke # rate speech editor The device includes a pre-emphasis filter that processes the original input voice signal 5ν (/ 7) into a pre-emphasis voice signal 忒 / 〇; a ^ second-order adaptive estimator, which uses an adaptive code This (Codeb00k) obtains a period parameter and the gain parameter through the least square difference, and the gain parameter is sum 并. For the benefit, the subtractor 'kills the above-mentioned pre-strong speech signal 〆々) to eliminate the above period parameter and Gain parameter to get an estimated signal? („) To get a residual value signal X / 7); a discrete cosine converter (DCT) receives the residual value signal of the above subtractor; ρ (Λ), takes an input frame every (# -point, according to Time domain overlap #, to establish the point conversion sound box, and then converted to #point DCT coefficient 4 left]; an auditory characteristic DCT coefficient picker, receives the aforementioned point DCT coefficient 4 magic, according to the human ear hearing characteristics and human frequency Variation of hearing tolerance, statistical characteristics of DCT coefficients, the coefficients of the points of view are divided into five different analysis areas in the frequency domain ^, and the absolute values of # coefficients are selected to form a DCT coefficient towards ^ '1 = {尤, / 2, _, form a base rate (3 fat 6 1 ^ 1 ^) encoding and high rate (High Rate) DCT coefficients, a DCT quantizer, the selected base rate above] ^ coefficient vector, weighted vector T ^ ector Quantization) to obtain the optimal codeword index parameter, 〇 and the optimal codeword quantization gain (Quantized Codebook GaiiOl, f). The DCT coefficients with a high rate are selected to use the quantized coding base. Rate DCT coefficient as a reference for scalar quantization (Scalar Quantization); The device integrates the above-mentioned period parameters, production and period gain parameters, such as the sum, DCT codeword index 胂, DCT quantization gain & and the quantization of the f DCT coefficient added due to the high rate to form a compressed data stream. Multi-rate speech compression;-A parametric demultiplexer that separates the above-mentioned compressed data stream into period parameters / * and period, now several Rong Heyu, DCT codeword index must be ic, # DCT positive and negative values, DCT quantization gain &, and the quantification of the coefficients given by the high rate, etc. Page 24 2002 · 06. 28.024 Agriculture 1 ^ 110061% Hold 1 ^^ Correction Wai · »♦ —- **** Lingyun-... ... ——— ππ 殆 first solution shrinking; truncated ft chord conversion composer, the DCT codewords solved by the above-mentioned parameter demultiplexer 32 2 leaching, more DCT positive and negative values, and DCT quantization gain & First, perform the I-point dct construction and then inverse quantize the additional DCT coefficients, and the remaining coefficients are reconstructed with zero to complete the reconstruction of the DCT coefficients. The cosine transform (IDCT) converter will reconstruct the pcT coefficients of the above-mentioned reconstruction points. Converted to the reconstructed sound box residual value signal, with # point before and # point after the previous heavy edge sound box. (KS0-point residual value signal; f-th order ^ estimation generator, the period parameter # and the period gain parameter solved by the above-mentioned parameter demultiplexer are summed to generate an estimator identical to the encoding end. ^ 3 above the reconstructed sound frame residual value signal and the estimated signal? (W) query to reconstruct the pre-strong speech signal); most waver #de (W 'processes the above-mentioned reconstructed pre-strong speech signal) to 2. ^ 3 | Build ^ tone signal%) Complete the decompression of the voice. One of the audio-visual characteristics described above is used for special emphasis; the coefficient coefficient of the [wave device H), the signal of Γ frame 5 °, is calculated by the household linear predictor (LPC)%, 7 = I 2, ···, / >}, where: rz H pre (?)-Σ a, a-Σ a / 'factory, which is a kind of “listening marriage” used in 歹 1 ^^ shall be The second-order adaptive estimator uses the codeword gain Pi corresponding to two consecutive adaptive vector vectors, 卩 2 'to generate an predictor. "The formula is as follows: where repulsion / ± 1, and the best soil parameter 2 Wood: ^ Λ, 2 >. The best base period parameter, = / *-1 as ϋ ^ ΙΙΙ page 25 202.0.28.025. 025: Conversely W = / * + 1 as (Shen Guang Yixi 2 < ^ SC ± +} 1 c > , c. ··, c7Tc7 > (sc, y ^ (sc,) A · He Fei ·, (: · * Long time squared to get the least square error _ 1 Voice Listening—Miscellaneous Xiao Yujia ft 5 — 5 · Voice Hearing characteristics: Used for the fun face TT p 7 qif ', where the coefficients of the DCT quantizer are outstanding in the frequency domain. HD: E is divided into five different analysis regions, which are 250Hz in two regions, 500Hz in C region, and iKHz in D region. , E area is 2KHz and other bandwidths. Ίίϊίτ ^ 数 ^ B, G, D The absolute value of each of the four ship domains is selected to dare to increase the DCT coefficient value and encode it with a gain vector; at a low rate "⑽Rate", add the code A, B An additional point of DCT coefficients in each of the two regions; in the middle rate (such as ⑹, the code C is added, and an additional point of dct coefficients in the two regions of D; in the case of HighRate), the code is increased based on the intermediate rate. 4 points in the e area j) ct coefficients 丄,) · k Application for the scope of patent application 丄 jL A sound auditory characteristic is used in the green 佶 turn 3Γ variable rate speech comprehension marquee when its input frame Domain (H ^ t When taking overlapping meeting points to create a #point conversion frame, a frame window function (FrameWind〇wFuncti () n) = 1, Μ < η <Ν-ΜΛ; »{/?] = (2/7 + 1 ) / 2 ^ 0 ^ / 7 ^; n {n] =:-\} / 2Μ, η 幺. 1. 7. · One of the items described in the first scope of the patent application ¥ The sound and hearing characteristics are used for Zi Ke's dual-rate speech coding solution Ma Ying, & DCT quantizer, the vector codebook with 4 dimensions and 4 vectors, in which 5 bits are used to index the Markov code word, and the gain of the code word size is 4 bits. The analysis is quantified with equal probability. 8. As described in item L of the patent application, one of the sound and hearing characteristics is used in Lubu ^ Variable Rate Speech Compilation Map, DCT quantizer i coefficient, and he applied for a patent In the coding architecture described in item 7 of the scope, additional areas A and B are added to select the additional additional coding DCT coefficients based on the auditory occlusion effect as a reference. And 7194 ___ Case No. 89110061 6. After normalizing the maximum DCT value in the patent application domain, quantify it in 3 digits scalar 丄 9. One of the sound and hearing characteristics described in the first patent application scope is used for variable speed Speech encoder, in which the coefficients of the DCT quantizer are added to the encoding architecture described in item 8 of the scope of the patent application for medium-rate encoding, and C and D are added to select the additional encoding point based on the auditory occlusion effect. DCT coefficient, normalized by the largest DCT value in the area, and quantized in 3 bits scalar 丄 10. A sound as described in item 1 of the patent scope is used for 特 to convert variable rate speech The codec, in which the number of DCT quantizers, is encoded at the high rate based on the encoding framework described in the patent application No. 9, and is additionally added in the E area according to the auditory occlusion effect, and is extracted in 3 bits. The positions of the 4 DCT coefficients are coded as shown in Table 1: Table 1 Position Index Position 1 Area E _000 32, 38, 44, 50 | 001_I 34, 40, 46, 52 010 36, 42, 48, 54 39, 45, 51 , 57 ~ ϊδο ^ ------- 41,47, 53, 59 101 43,49, 55, 61 --Τ ^ - 45,51,56, 6211148, 55, 58, 60 Page 27 2002.06.28.027