1253057 九、發明說明： 【發明所屬之技術領域】 本發明係關於-種編石馬向量(C0de_vect0r)的搜尋系統及其方 法」！寺別是關於-種於-言吾音編碼器(Speech enc〇der)中搜尋二語 音訊號(Speech signal)之一標的編碼向量的搜尋系統及其方法。 【先前技術】 4知適應性夕重速率(Adaptive multi-rate，AMR)數位語音編 碼標準是由第三代行動通訊夥伴合作計晝(Third获收加籲 partnership project ’ 3GPP)所訂定。根據 AMR 標準規格 3Gpp TS 26.090，總共有八種低位元語音編碼模式，分別為 12.2kbps、10.2kbps、7.95kbps、7.40kbps、6.70kbps、 5.90kbps、5.15kbps、以及 4.75kbps。AMR 標準採用的核心 技術為代數激式線性預測（Algebraic c〇de_excited linear_ prediction，ACELP)。 請參閱圖一，圖一為習知ACELP語音編碼器1〇的示意 圖。ACELP語音編碼器1〇包含一前置處理器（pre一 processor^、一線性預估分析器(Unear predict肪咖撕)14、一參 適應性編碼;4搜寻态(Adaptive codebook search)16以及一代數編 碼>春搜哥态(Algebraic codebook search)l8。前置處理器12包含 一高頻濾波器20。首先，一語音訊號s(n)輸入前置處理器 12，並藉由前置處理器12中的高頻濾波器2〇來濾掉語音 訊號s(n)的直流成份。之後，語音訊號s(n)再經過線性預估 分析态14，產生一激發訊號(Excitation signal) X⑻。激發訊號 X⑻可經由一週期性激發訊號以及一代數碼激發訊號合成 產生。激發訊號X⑻先經過適應性編碼薄搜尋器16，得到週期 性激發訊號，並且經由代數編碼薄搜尋器18，產生代數碼 激發訊號。 1253057 於ACELP語音編碼器ι〇中 尋找-較佳的編碼向量㈣及其㈣^數辦_尋器18係用以 與標的訊號(Target signal) &的均二* & ’使^寻合成的語音訊號 可經由下列公式一計算得出·· σ、差為最小。均方誤差£k (X2 —, 公式一： f二’公式—中的％係代表位於代數編碼薄中第k彳_帥旦。 大的決策分數(DedSi〇nscore)Ak。m八季 = 的、扁碼向里可產生較 由下列公式二計算得出： &^D~e)Ak可經 公式二：Ak=fl=^^ , 八中d h x2為;^的成號Χ2與線性 h(n)^a ^decorrelation function)， 應h(n)的自相關函數(Aut〇_c〇rrelati〇n化沉丨沁幻。 由於代#丨_驗尋If l8佔了整個ACELp語*編碼器1〇 ^部分的计异夏，所以許多文獻皆提出有效率的編碼向量搜 尋方法，致力於簡化代數編碼薄搜尋的計算複雜度(c〇mputing complexity)，並且提升語音品質（Speech quamy)，例如：美 國專利公告號第5,701，392號、美國專利公告號第6,714,9〇7號、 Hochong Park, ^Efficient Codebook Search Method for EVRC Speech1253057 IX. DESCRIPTION OF THE INVENTION: TECHNICAL FIELD OF THE INVENTION The present invention relates to a search system and method for a stone horse vector (C0de_vect0r)! The temple is a search system and method for searching for a coded vector of one of the speech signals in the speech enc〇der (Speech enc〇der). [Prior Art] 4 The Adaptive Multi-rate (AMR) digital speech coding standard is defined by the third generation mobile communication partner cooperation plan (Third received and added partnership project ' 3GPP). According to the AMR standard specification 3Gpp TS 26.090, there are a total of eight low-order speech coding modes, 12.2 kbps, 10.2 kbps, 7.95 kbps, 7.40 kbps, 6.70 kbps, 5.90 kbps, 5.15 kbps, and 4.75 kbps. The core technology used in the AMR standard is Algebraic c〇de_excited linear_prediction (ACELP). Please refer to FIG. 1. FIG. 1 is a schematic diagram of a conventional ACELP speech coder. The ACELP speech encoder 1 includes a pre-processor (pre-processor^, a linear predictor (Unear predict), a parametric code; 4 (adaptive codebook search) 16 and a generation The number encoding > Algebraic codebook search 18. The pre-processor 12 includes a high frequency filter 20. First, a voice signal s(n) is input to the pre-processor 12 and processed by pre-processing. The high frequency filter 2 in the device 12 filters out the DC component of the speech signal s(n). The speech signal s(n) is then subjected to a linear prediction analysis state 14 to generate an excitation signal X(8). The excitation signal X(8) can be generated through a periodic excitation signal and a generation of digital excitation signals. The excitation signal X(8) is first passed through the adaptive code search engine 16 to obtain a periodic excitation signal, and the algebraic code search engine 18 is used to generate the digital excitation. 1253057 Look for the ACELP speech coder ι〇 - the preferred coding vector (4) and its (4) _ _ _ _ 18 is used to match the target signal (Target signal && 'amplifier ^ Synthetic language The signal can be calculated by the following formula: · σ, the difference is the smallest. The mean square error £k (X2 —, the formula 1: f 2 'formula—is the part of the algebraic codebook. The large decision score (DedSi〇nscore) Ak.m8 season =, the flat code inward can be calculated by the following formula 2: &^D~e)Ak can be calculated by formula 2: Ak=fl= ^^ , 八中 dh x2 is; ^ is the number Χ 2 and linear h(n)^a ^decorrelation function), should be h(n) autocorrelation function (Aut〇_c〇rrelati〇n 丨沁 丨沁 丨沁Since the generation #验_检解If l8 occupies the entire ACELp language* encoder 1〇^ part of the different summer, so many documents have proposed efficient coding vector search method, which is devoted to simplifying the computational complexity of algebraic codebook search. (c〇mputing complexity), and the improvement of speech quality (Speech quamy), for example, U.S. Patent No. 5,701,392, U.S. Patent No. 6,714,9,7, Hochong Park, ^Efficient Codebook Search Method for EVRC Speech
Codec’’，IEEE Signel Processing Letters, vol· 7，no. 1，2000、 Hochong Park，Younchang Choi and Doyoon Lee，“Efficient Codebook Search Method for ACELP Speech Codecs，，，IEEE， 2002、…專。评估代數編碼薄搜尋的效能指標(perf〇rmance measurement)包含計算複雜度與語音品質。在計算複雜度方 面’可依據處理器計算ACELP語音編碼器10所需的時間來評 估。在语音品質方面’可依據語音品質聽覺評價（perceptuai evaluation of speech quality，PESQ)值來評估。PESQ 是由 1253057 國際通訊協會(ITU-T)於規格ITU-丁 R862中所訂定。PES 用客觀性評量的聽覺模型估計主觀性評量的M〇s(Mean opimon SCOre)值。PESQ值的範圍係介於·〇·5至4 $之 PESQ值越南，表示語音品質越好。 、门3GPP所訂定的AMR標準，其採用的代數編碼薄搜尋方法 為取珠樹狀搜尋法(Depth-first tree search)。該方法詳细描述於 AMR標準規格3GPP TS 26·09〇與美國裳 5,701，392 號。 氣弟 請參閱圖二，圖二為根據AMR標準的122kbps模式 中，一編碼向量的脈衝位置分佈圖。代數編碼薄中每一個編^ 向量係分別定義40個脈衝位置(pulse position)並且包含1〇個脈衝 (Pulse)，其中脈衝位置以n表示，ρο，〗，…，39，脈衝以&表 示，卜〇，1”..，9。如圖二所示，10個脈衝係平均分佈於5個執道 (Tt，t=0，l，...，4)的40個位置中。因此，每一個脈衝可能出現 在其相對應執道的8個位置上。以脈衝P()為例，ρ〇可能出 現在轨道T〇的n=〇,5，10，15,20,25,30,35等8個位置上。代數 編碼薄搜尋方法即由40個位置中搜尋出1〇個脈衝，以構成較 佳的編碼向量ck，使得決策分數Ak較大，亦即合成 與標的訊號的均方誤較小。 #u 4請參閱圖三，圖三為習知技術根據八^標準之最深樹狀 搜尋法之流程圖。以12.2kbps模式為例，根據AMR標準，最 深樹狀搜尋法包含下列步驟： 步驟S100 :開始搜尋。 •步驟S102 :經由下列公式三，計算每個脈衝位置的命中函數 (Hit function) b(n)； 1253057 公式三：6⑻ resLTP (η) d{n) 0，1，2,·_·..，39 其中.reSLTP⑻為脈衝位置n白勺長時預測剩餘 prediction residual)訊號’ d⑻為脈衝位置n的標的訊 線性預估濾波器脈衝響應h(n)的相關函數； a h、 置； 步驟S104 :設親衝Pq的脈衝位置於最大b(顺的脈衝位 步驟S106 :設定脈衝Pl的脈雜置於脈衝 的其他四個軌道中，具有最大b(n)值的脈衝位斤在執運以外 步驟S108 :於脈衝Pl的次一軌道與次二軌道 J Ak，搜尋脈衝P2* p3的脈衝位置，舉例來說， ^，則次-執道即為執道％且次二執道即為 依此類推； 1 M下步驟 步驟S11G :於脈衝P3的次—執道與次二 數Ak，搜尋脈衝p4與p ; 队錄决朿分 依據決策分 步驟S112 :於脈衝p5的次—軌道與次二執道， 數Ak，搜尋脈衝？6與？7 ; H114 :於脈衝p7的次一軌道與次二軌道 數Ak，搜尋脈衝？8與！>9 ; 像决朿分 步驟S1:，：檢查是否已翻—預定搜尋:欠數，若是 S118，若否，則回到步驟sl〇6 ; 、進行 步驟S118 :依據所搜私每—組脈衝(&，···⑹ 向1的決鼓分數，ip遥异杜a , 项4的編碼 向ί ^擇取佳的10個脈衝位置以決定一標的編碼 8 1253057 步驟S120 :結束搜尋。 如上所述，若預定搜尋次數定為四次，則AMR最深樹狀搜 尋法共需搜尋4*(8*8+8*8+8*8+8*8)=1024次，以決定編碼向 量。 請參閱圖四，圖四為習知技術之脈衝取代搜尋法(Pulse replacement search)之流程圖。脈衝取代搜尋法配合最深樹狀搜尋 法可以改善編碼音質。脈衝取代搜尋法包含下列步驟： 步驟S200 ·開始搜尋。 步驟S202 :經由上述最深樹狀搜尋法取得一預選編碼向量， 並計算預選編碼向量之決策分數； 步驟S204 :計算預選編碼向量中每一個脈衝位置的貢獻分數 (Contribution score)； 步驟S206 :選取預選編碼向量中，貢獻分數最低的脈衝位置 以及其執道； :由該執道其他脈衝位置中，選擇―㈣脈衝位置 貝f分數的脈衝位置，該候選脈衝位置所構成的-候 一扁碼向1相較於錄道其他脈衝位置具錄高決策分數； 石弓向:確認是否該候選編碼向量之決策分數低於預選編 步前預選編碼向量為標的編碼向量，並進行 /驟幻16,若否，則進行步驟S212; 置，=候脈衝位置取代最«獻分數的脈衝位 又祈候廷編碼向1為預選編碼向量； 數，位置的取代缝—舦取代次 疋貝J進仃步驟S216，若否，則回到步驟S204 ; 1253057 步驟S216 :結束搜尋。 請參閱圖五，圖五為習知技術之次 ^ (Sub-codebook search)之流程圖。根據美國、、兩竭溥搜尋法 6,714,907號，次代數編碼薄搜尋法包含下列步驟利公告號第 步驟S300 :開始搜尋。 步驟S302 :經由最深樹狀搜尋法取得第一 > 最佳預選編碼向量，並計算預選編碼向量之決策分^戈數編碼薄的 步驟S304 :經由最深樹狀搜尋法取得下— 最佳候選編碼向量，並計算候選編碼向量之決策分數編碼薄的 步驟S3〇6 :比較預選編碼向量之決策分數盘候 決朿分數，並選出最佳的決策分數以及編碼向量；、^馬向量之 步驟S308 :確認是否為最後的次代數編瑪，若 驟S310，若否，則回到步驟幻⑽； 則進行步 步驟S310 :對最佳編石馬薄的編碼向 搜尋，以決定最後的編碼向量； 錢仃上叙脈衝置換法 步驟S312 :結束搜尋。 哭大數編碼薄搜尋佔了整個概…_ 口口大邛刀的计异置。以AMR 12 2kb 口口曰、、扁碼 搜尋器使用最深樹狀搜尋法佔Ae ^^為例，代數編碼薄 約的計算量，總共了 =4P =音1爲碼器整體語音編碼 佳編碼語音品質，^雖可得到較 1253057 -語音目的在於提供—種於—語音編碼器中搜尋 為虎之柄向1的搜尋綠及其綠，啸虹述問題。 【發明内容】 語音目的在於提供—種搜尋系統及其方法，用來於- enc〇der)^^^-^i-mf#u(Speech signal)^-:質、，。馬向里（arget eQde_vee㈣，以降低計算複雜度並確保語音 f卢之二法係用以於—語音編碼器中’搜尋-語音訊 Ϊί旦語音訊號包含複數個編碼向量，該等編 脈衝係可分配在位於該編碼向量的該等脈衝位 衝位置係分佈於複數個執道(Track)上。本發明之搜 咢方法包含下列步驟·· ^ K仅 ⑻計算每-脈衝位置的—相對應命中函數_⑽魅）； (b) 根據每一轨道中的脈衝位置個數與脈衝個數，決定每一轨 道中的複數組脈衝組合(Pulse c⑽binatiQn); (c) 根據每-脈衝位置的相對應命中函數，計算每一脈衝組合馨 的一相對應組合命中函數(Combination hit functior〇 ; ⑻根據每-脈衝組合的相對應組合命中函數大小，依序排列 每-執暹巾的脈衝組合，選縣—執道巾具有最大組合命中函數 的脈衝組合，以作為一相對應預選脈衝組合(Default pulse combination)，並將其他脈衝組合依大小次序排列成一次序串列； ⑻根據每-軌道中的預選脈衝組合，形成—預選編碼向量 (Default code-vector)，並計算預選編碼向量的一決策分數 (Decision score)； 1^53057 軌道的預選脈衝組合，一以⑨蚪替代候選脈衝組合相對應 vector), code- 據候選編碼向量以及預選編碼向量㈣t八數，、隹> 、、扁馬向置更新程序，以決定出標的編碼向量的决朿刀數進仃一 佳編二量量搜尋方法，不僅可減少搜尋較 的決策分而复雜度，並且可產生較大 保語音品質。°曰5fl號與編碼語音訊號的誤差，進而確 2 =憂點與精神可以藉由以下的發明詳述及所附 式得到進一步的瞭解 【實施方式】 圖Codec'', IEEE Signel Processing Letters, vol. 7, no. 1, 2000, Hochong Park, Younchang Choi and Doyoon Lee, "Efficient Codebook Search Method for ACELP Speech Codecs,, IEEE, 2002, .... Evaluation of algebraic coding The perf〇rmance measurement of thin search includes computational complexity and speech quality. In terms of computational complexity, it can be evaluated according to the time required by the processor to calculate the ACELP speech coder 10. In terms of speech quality, it can be based on speech. The perceptuai evaluation of speech quality (PESQ) value is evaluated. The PESQ is defined by the 1253057 International Telecommunications Association (ITU-T) in the specification ITU-R R862. PES uses subjective assessment of the auditory model to estimate subjective The M〇s (Mean opimon SCOre) value of the sexual evaluation. The range of PESQ values is in the PESQ value of 〇·5 to 4 $ Vietnam, indicating that the voice quality is better. The AMR standard set by the 3GPP, The algebraic codebook search method used is the Depth-first tree search method. The method is described in detail in the AMR standard specification 3GPP TS 26·09〇 and the US Supreme 5, No. 701, 392. Please refer to Figure 2 for the gas brother. Figure 2 shows the pulse position distribution of a code vector in the 122 kbps mode according to the AMR standard. Each of the algebraic codebooks defines 40 pulse positions (pulse). Position) and contains 1 pulse (Pulse), where the pulse position is denoted by n, ρο, 〗 〖, 39, and the pulse is represented by & 〇, 1".., 9. As shown in Figure 2, the 10 pulse trains are evenly distributed in 40 locations of 5 orbits (Tt, t = 0, 1, ..., 4). Therefore, each pulse may appear in eight positions of its corresponding way. Taking the pulse P() as an example, ρ〇 may appear at 8 positions such as n=〇, 5, 10, 15, 20, 25, 30, 35 of the track T〇. The algebraic codebook search method searches for one pulse from 40 locations to form a better code vector ck, which makes the decision score Ak larger, that is, the mean square error of the synthesized and the target signal is smaller. #u 4 Please refer to Figure 3, which is a flow chart of the deepest tree search method according to the eight standard. Taking the 12.2 kbps mode as an example, according to the AMR standard, the deepest tree search method includes the following steps: Step S100: Start searching. • Step S102: Calculate the hit function b(n) for each pulse position via Equation 3 below; 1253057 Equation 3: 6(8) resLTP (η) d{n) 0,1,2,·_·.. , 39 where .reSLTP(8) is the long-term prediction residual residual residual of the pulse position n. The signal 'd(8) is the correlation function of the target linear prediction filter impulse response h(n) of the pulse position n; ah, set; step S104: Let the pulse position of the Pq pin be at the maximum b (the punctual pulse bit step S106: the pulse of the set pulse P1 is placed in the other four tracks of the pulse, and the pulse bit having the largest b(n) value is outside the step of the execution. S108: searching for the pulse position of the pulse P2*p3 in the next track and the second track J Ak of the pulse P1, for example, ^, then the second-executive is the % of the way and the second is the right way. Analogous; 1 M next step S11G: in the pulse P3 sub-executor and the second binary Ak, search for the pulse p4 and p; the team is determined according to the decision step S112: the second-track and the second in the pulse p5 Execution, number Ak, search pulse? 6 and ?7; H114: the second and second orbital number Ak of the pulse p7, The search pulse 8 and !>9; the decision step S1:,: check whether it has been turned over - the predetermined search: the number of underruns, if it is S118, if not, then return to step sl6; The search for each group of pulses (&,···(6) to the drum score of 1 , ip remote d a, the code of item 4 to ί ^ select a good 10 pulse position to determine a target code 8 1253057 Step S120: Ending the search. As described above, if the predetermined number of search times is set to four times, the AMR deepest tree search method needs to search for 4* (8*8+8*8+8*8+8*8)=1024. Second, to determine the coding vector. Please refer to Figure 4, Figure 4 is a flow chart of the pulse replacement search method of the prior art. The pulse substitution search method and the deepest tree search method can improve the coding sound quality. The method includes the following steps: Step S200: Start searching. Step S202: Obtain a preselected coding vector through the deepest tree search method, and calculate a decision score of the preselected coding vector; Step S204: Calculate the contribution of each pulse position in the preselected coding vector Contribution score; step S206 : selecting the pulse position with the lowest contribution score and its obedience in the preselected coding vector; : selecting the pulse position of the (four) pulse position and the f-score from the other pulse positions of the obedience, the candidate pulse position The flat code has a higher decision score than the other pulse positions of the recorded track; the arch bow direction: confirm whether the decision score of the candidate code vector is lower than the pre-selected pre-selected code vector as the target code vector, and perform/fantasy 16 If not, proceed to step S212; set, = wait pulse position to replace the most fractional pulse bit and pray to encode 1 to the preselected coding vector; number, position of the replacement seam - 舦 replace the secondary mussel J Step S216, if no, return to step S204; 1253057 Step S216: End the search. Please refer to Figure 5, which is a flow chart of the sub-codebook search. According to the United States, and the two exhaustive search methods 6,714,907, the subalgebraic codebook search method includes the following steps: Step number S300: Start searching. Step S302: Step S304 of obtaining a first > optimal preselected coding vector via the deepest tree search method, and calculating a decision classification code of the preselected coding vector: obtaining the next best candidate coding through the deepest tree search method a vector, and a step S3 〇 6 of calculating the decision score code of the candidate code vector: comparing the decision score of the preselected code vector to the score, and selecting the best decision score and the code vector; step S308 of the horse vector: Confirm whether it is the last sub-algebra, if it is S310, if not, then return to step magic (10); then proceed to step S310: search for the best coded horse thin to determine the final coding vector; Step S312: Ending the search. Cry large number coded thin search accounted for the entire ... _ mouth slashing the meter. Take AMR 12 2kb port 曰, flat code searcher using the deepest tree search method to account for Ae ^ ^ as an example, algebraic coding thin calculation amount, a total of = 4P = sound 1 is the coder overall speech coding good coded speech Quality, ^ can be obtained more than 1253057 - the purpose of the voice is to provide - in the vocal encoder search for the tiger's handle to 1 search for green and its green, Xiaohong. SUMMARY OF THE INVENTION The purpose of speech is to provide a search system and method thereof for use in - enc〇der)^^^-^i-mf#u(Speech signal)^-: quality,,. Ma Xiangli (arget eQde_vee (4), to reduce the computational complexity and ensure that the voice of the two systems is used in the - vocal encoder 'search - voice signal Ϊ 旦 语音 voice signal contains a plurality of code vectors, the pulse system can The pulse position positions assigned to the code vector are distributed over a plurality of tracks. The search method of the present invention includes the following steps: · K only (8) calculating the per-pulse position - corresponding hit Function _(10) charm); (b) Determine the complex array pulse combination in each track according to the number of pulse positions and the number of pulses in each track (Pulse c(10)binatiQn); (c) Corresponding to the position of each pulse Hit function, calculate a corresponding combination hit function for each pulse combination (Combination hit functior; (8) according to the size of the corresponding combination hit function of each-pulse combination, sequentially arrange the pulse combination of each-shouldered Siam, select the county - the towel has the pulse combination of the largest combined hit function as a corresponding default pulse combination, and the other pulse combinations are arranged in a size order. (8) according to the preselected pulse combination in each-track, form a pre-selected code vector (Default code-vector), and calculate a decision score of the pre-selected code vector; 1^53057 orbit preselected pulse combination, One is to replace the candidate pulse combination with the vector), the code-based candidate coding vector and the pre-selected coding vector (four) t eight, 隹>, and the flat horse orientation update procedure to determine the standard coding vector. The number of the two-quantity search method can not only reduce the complexity of the search decision, but also produce a larger voice quality. °曰5fl and the error of the encoded speech signal, and then 2 = worry and spirit can be further understood by the following detailed description of the invention and the accompanying formula [embodiment]
依夂六丄圖六為根據最深樹狀搜尋法，於各軌道中， 丄—之吁中函數值大小排列的脈衝出現機率分佈圖。於 H ’貫驗語音的長度共616個語音訊框(sPeech fr_)，亦 7 "且母個軌道各出現4928個脈衝。如圖六所示， 罢巾當脈衝位置的命巾函數值越大，脈衝出現在此脈衝位 ^的機率也越大。以軌道TG為例，脈衝位置的命中函數值最大 日寸’脈衝出現的機率也最大(41·6%)。其次，隨脈衝位置的命中 函數值減少，脈衝出現的機率也遞減。因此，本發明乃是根據各 脈衝位置的命+函數，來蚊鎌触脸合之組合命中函數， 預測較佳的脈衝組合序列，以降低搜尋時的計算複雜度。 請參閱圖七，圖七為本發明之編碼向量的搜尋系統3〇示意 圖。本發明之搜尋系統3〇係用以於一語音編碼器（Speech encoder)(圖七中未顯示）中，搜尋一語音訊號物沈也signai)之一標 12 1253057 code^vector) 〇 二丄该等編碼向量分別疋義出複數個脈衝位置(pulse p〇siti〇n)並且 =複數個脈衝(Pulse)，每-脈衝係可分配在位於該編碼向 j脈衝位制，該雜衝位置係分佈於複數働道(τ她)上。、 搜哥糸統30包含一第一裝置(Device)32、一第二裝置34、二 裝置36、一第四裝置38、一第五裝置 — 第七裝置44。 币，、衣直42以及_ 第衣置32可以疋一處理器（processor)或一計算哭 ㈣，主要是設料㈣計算每—_似的—相對^ ^函數，―）。第u 34可以是—處理器或一控 forf) ’主要是料為轉每—鏡巾雜衝位置個數盘脈 衝，數·’用j決定每—執道中的複數組脈衝組 comb融cm)。弟三裝置36可叹—處㈣或—計算器」 設^根據每-脈衝位置的相對應命中函數，用以計算每一脈g 、，且&的-相對應組合命中函數(c〇mbinati〇n汕細 置38可以是—處理器或-控制器，主要是設計= ί 脈？組合，，相對應組合命中函數大小，用以依序排列每一該等 執迢中的該雜触合’選縣—鮮 ^中 函數的，衝組合：作為一相對應預選脈衝= 衝組合依大小次序排顺—次序串列。 弟五衣置40可乂疋一處理器或一算哭，主 -執道中的預選脈衝組合，用㈣/…^疋以為根據母 ^ ^ 用以形成一預4編碼向量(Default C〇de_VeCt〇r)，亚汁鼻預選編碼向量的一決策分數(ILcision score)。第六裝置42可以是一處理$ 管 （ 選擇次一脈衝組合為一 要是設計為 1 . . X v 献士侯 & 脈衝、、且否（Candidate pulse ，maio^i) ’ θ日守日代候選脈衝組合相對應軌道的預 合，而形成一候選編碼向|(r 」 、、訂、、 編石弓向量的-料ee㈣’並計算候選 馬：t 刀數。弟七裝置44可以是-處理哭ϋ制 器，主要綱為根據候選編碼向量以及預選編;向;以 13 1253057 數，用以決定丨標的編碼向量。 請參閱圖八，圖八发θ 置44進一步包含^圖丄中之第七裝置44*意圖。第七裝 第三模組50。第-‘ tf^fe)46、；：第二模組48以及一 設計為用以判斷候選έ扁！ 處理&或—控制器，主要是 設計為用以更新候選編—it、處理益或―控制器，主要是 以是-處理器或一控制器：::三模組50可 時，搜尋系統3G t目搜尋結果符合默搜尋條件 完成搜尋。 則的預遠編碼向量為標的編碼向量，並i 衝位ίί:ΐϊ數ΪΙΗ圖::(A)以及圖十-⑼’圖九為各脈 的第一 所決定的預選脈衝組合示意圖。圖十 施例中，組合命巾函㈣相對應之 1 ^ 相對應命中函數的總和。 丫々卿X位置的 根據本發明之第-實施例，以AMR 12.2kbps模式 ? ’看，位置分佈係如圖二所示。減上述本發明之編碼向 量搜哥土統30 語音編碼器接收一語音訊號時，搜尋系 統30冒開，搜尋該語音訊號的編碼向量。以下將以一^固 標的編碼向量的搜尋過程配合圖七以及圖八做說明。 第-裝置32先計算該編碼向量中每個脈衝位置的命中 b(n) ’如®九所tf。帛二裝置34根據各減之脈触置盘脈衝 個數’決定所有執道之脈衝組合。於AMR 12 2Kbps模式;，^ 個執道皆具有兩個脈衝在八個可能脈衝位置上(可重複），因此每 1253057 ，軌道的脈衝位置組合共有，=36種可能。例如， 對瘅人巾7|，呌管）弟一衣置36根據母一脈衝位置的相 脈衝组合的—相對應紐合料函數。舉 的命中函數值標示於各脈衝组人的 θ 不各脈衝組合 據各脈衝組合之組合命中函數 =]弟四衣置38根 時，各執道中最大組合命中母；:執道中的脈衝組合 (25,25)、執道Tl _ Γ/J^f之脈衝組合分別為執道T0的▲ 軌道丁4的(19 19)。上，^= 2的(7,7)、執道13的(33,33)以及^ 一⑷所矛:。組合即為預選脈衝組合，如圖十 小，排成一次序；歹r，：欠依組合命中函數大 (1,7,7,16,19,19,25 25 33 33)，構成候選編碼向量According to the six-figure map six, according to the deepest tree-like search method, in each track, the probability distribution pattern of the pulse of the function value size in the 丄- 吁 之. A total of 616 voice frames (sPeech fr_), also 7 " and a total of 4928 pulses per parent track. As shown in Figure 6, the greater the value of the life towel function at the pulse position, the greater the probability that the pulse will appear in this pulse position ^. Taking the orbit TG as an example, the probability of the hit value of the pulse position is the largest (41.6%). Second, as the value of the hit function decreases with the pulse position, the probability of a pulse appearing decreases. Therefore, the present invention predicts a better pulse combination sequence based on the life + function of each pulse position, and combines the hit function of the mosquito face to predict the computational complexity of the search. Please refer to FIG. 7. FIG. 7 is a schematic diagram of a search system of the code vector of the present invention. The search system 3 of the present invention is used in a speech encoder (not shown in FIG. 7) to search for a voice signal (signa) and a signai) 12 1253057 code^vector) The equal coding vectors respectively represent a plurality of pulse positions (pulse p〇siti〇n) and = a plurality of pulses (Pulse), and each pulse system can be allocated in the code to j pulse position, and the miscellaneous position distribution On the plural martyrdom (τ she). The search system 30 includes a first device 32, a second device 34, two devices 36, a fourth device 38, and a fifth device-seven device 44. The coin, the clothing straight 42 and the _ the first clothing 32 can be a processor or a computing cry (four), mainly the material (four) calculate each - like - relative ^ ^ function, -). The u 34 can be - processor or a control forf) 'mainly is expected to turn each - the number of mirrors of the mirror is a number of disk pulses, the number of 'use j to determine the complex array of pulses in each channel. . The third device 36 sighs - (4) or - the calculator" sets the corresponding hit function according to the position of each pulse to calculate each pulse g, and the corresponding combination hit function of the pulse (c〇mbinati 〇n汕 置 38 may be a processor or a controller, mainly a design = ί pulse combination, corresponding to the combined hit function size, for sequentially arranging the heterogeneous touch in each of the stubs 'Select county-fresh ^ function, rush combination: as a corresponding pre-selection pulse = rush combination according to the size order shun-order string. Brother five clothes set 40 can be a processor or a cry, the main - The pre-selected pulse combination in the obedience is determined by (4)/...^ to form a pre-4 coding vector (Default C〇de_VeCt〇r) according to the parent ^^, and a decision score (ILcision score) of the juice vector preselected coding vector. The sixth device 42 can be a processing $tube (selecting the next pulse combination as one if the design is 1. . X v 士士侯 & pulse, and no (Candidate pulse, maio^i) ' θ 日守日代The candidate pulse combination corresponds to the pre-composition of the orbit, and forms a candidate code to |(r ′, Order, edit the stone bow vector - material ee (four) 'and calculate the candidate horse: t knife number. The brother seven device 44 can be - handle the crying controller, the main program is based on the candidate coding vector and pre-selected; direction; to 13 1253057 For determining the coding vector of the target. Referring to Figure 8, the eighth θ set 44 further includes the seventh device 44* in the figure. The seventh installed third module 50. -'tf^fe) 46;;: The second module 48 and a design for determining the candidate ! flat! Processing & or - controller, mainly designed to update the candidate -it, processing benefits or - controller, mainly Yes - processor or a controller::: three modules 50, the search system 3G t mesh search results meet the default search conditions to complete the search. Then the pre-far code vector is the target code vector, and i is bit ίί:ΐϊ The number map:: (A) and Figure 10 - (9) 'Figure 9 is the first determined pre-selected pulse combination diagram of each pulse. In the example of Figure 10, the combined life towel function (4) corresponds to the 1 ^ corresponding hit function The sum of the XQ position X according to the first embodiment of the present invention, with AMR 12.2 kb Ps mode? 'Look, the position distribution is as shown in Fig. 2. When the above-mentioned code vector search system 30 vocal encoder receives a voice signal, the search system 30 pops up and searches for the code vector of the voice signal. The following is a description of the search process of a fixed-coded vector with FIG. 7 and FIG. 8. The first device 32 first calculates the hit b(n) of each pulse position in the code vector, such as the nine nine tf. The device 34 determines the pulse combination of all the commands based on the number of pulses of each of the subtracted pulses. In the AMR 12 2Kbps mode; ^^ each channel has two pulses at eight possible pulse positions (repeatable), so every 1253057, the pulse position combination of the track is common, = 36 possibilities. For example, for the 巾人巾7|, 呌管)一衣36 is a function of the corresponding kinetics of the phase pulse combination according to the position of the mother-pulse. The value of the hit function is indicated by the θ of each pulse group. The combination of each pulse combination is based on the combined hit function of each pulse combination =] the third combination is 38, the largest combination hits the parent in each way; 25,25), the pulse combination of the command Tl _ Γ / J ^ f is the ▲ track D4 of the road T0 (19 19). On, ^= 2 (7,7), Ego 13 (33,33) and ^1 (4) spear:. The combination is a pre-selected pulse combination, as shown in Figure 10, arranged in a sequence; 歹r,: under-combination hit function is large (1,7,7,16,19,19,25 25 33 33), constitutes the candidate coding vector
Ac。接著，第七裝置’J中的m選^，量之決策分數_ 質；若候選編碼向量之決簠八赵^脈衝組合⑽以改善語音音 分數〜，則第七〜以4朿;：。=;編碼向量之決策 (1，16)取代預選脈衝組合（fUt以候選脈衝組合 (1，7,7，16，19，19,25,25,33,33)及更新決策八數、預遠編碼向量為 七裝置44中的第三模組5〇 U查為AC。最後，第 搜尋條件，若是，則以 ^ 果是否符合-預定 完成搜尋。 預&、、扁碼向置為標的編碼向量，並 1253057 於此實施例中，如圖十一所示，芒抽夺^ π 合(^18)θ，則停止搜尋，並取得較佳決策分脈= 注意的是，雖然本實施例的預定搜尋條件尋至。^ = 的最後一個脈衝組合為止，但為了降低搜尋時中 分脈衝組合。由圖六的實驗結果可以看 去一 命中函數值太低時，脈衝出現於該脈衝 ^ 置= 、、且口以即桃哥日守間。換句話說，本發明可進 界值(Threshold)，若-脈衝組合之組合命中二二™ 則將該脈衝組合排除於次序串列。本發 介值，則不參與排序。又例如，，床 衝組合’組合命中函數值較低的脈衝組合就不會列脈 此外，預定鮮餅可為—歡鮮缝或—取搜尋H ° 凊參閱圖十二、圖十三（A)以及圖十三 根據本發明的第二實施例，所有脈衝組合 意。立圖十三⑷為根據本發明的第二實施例 f 口 ^圖。圖十三⑻為根據本發明的第二實施例所決定 二實施例中，組合命中函數係相:應{ 的次據母一脈衝位置的相對應命中函數大小所決定 根據本發明之第二實施例，以AMR 12 2kbDS 盔 衝位置分佈係如圖二所示。第二實施例盘第一 的主要差異在於組合命中函數的定蠢 二貝轭例 8*〇Wn〇mb(n2))： #t 斜表不麵位置ni於賴道命中函數值的次序， I ： 16 1253057Ac. Next, m is selected in the seventh device 'J, the decision score of the quantity is _ quality; if the candidate coding vector is determined by the combination of the octaves (10) to improve the speech score ~, then the seventh to 4 朿; =; coding vector decision (1,16) replaces preselected pulse combination (fUt with candidate pulse combination (1,7,7,16,19,19,25,25,33,33) and update decision eight, pre-far The code vector is the third module 5〇U of the seven devices 44 is checked as AC. Finally, the search condition, if yes, is whether the search is completed or not. The pre-amplitude, the flat code is set to the target code. Vector, and 1253057. In this embodiment, as shown in FIG. 11 , when 芒 ^ ^ π (^18) θ, the search is stopped, and a better decision is made. Note that although the reservation of this embodiment is made The search condition is found until the last pulse combination of ^ =, but in order to reduce the mid-divided pulse combination in the search. It can be seen from the experimental result in Fig. 6 that when the value of a hit function is too low, the pulse appears in the pulse set = , In other words, the present invention can enter the threshold (Threshold), if the combination of the -pulse combination hits the second and second TM, the pulse combination is excluded from the sequence. , then do not participate in the sort. For another example, the bed punch combination 'combined hit function value lower pulse In addition, the combination will not be inline, in addition, the predetermined fresh cake may be - fresh seam or - search H ° 凊 see Figure 12, Figure 13 (A) and Figure 13 according to the second embodiment of the present invention, all pulses Figure 13 (4) is a diagram of a second embodiment according to the present invention. Figure 13 (8) is a second embodiment according to the second embodiment of the present invention, the combination hit function phase: According to the second embodiment of the present invention, the AMR 12 2 kb DS helmet position distribution system is shown in Fig. 2. The main difference of the first embodiment is that the first difference is that The order of the combined hit function is 8*〇Wn〇mb(n2)): #t The order of the non-surface position ni is in the order of the value of the hit function, I : 16 1253057
Kn^p-bCn2)。於此實施例中，命中函數值b(ni)或 組合命中函數值越小，次序越先。如圖十二所示，二（大， 上，脈衝組合依序為(25,25)、（25,30)、（〇,25)、 、（5 道丁0 脈衝組合(25,25)的組合命中函數值為〇、脈衝組合;^二其中 命中函數值為1，依此類推。因此，當第四裝置38根)== 組合之組合命中函數，依序排列每一執道中的脈衝钽人萨^ ' 道中最先位置命中函數之脈衝組合分別為軌道丁0的°^，口幸九 L的(1，1)、執道τ2的(7,7)、執道τ3的(33,33)、以及執、九， (一 19，19)。上述五個脈衝組合即為預選脈衝組合，如圖 4、 不。於此實施例中，其他脈衝組合依軌道位置依序 = 串列，如圖十三(Β)所示，該次序串列依序為(25,3 =序丨 (7,22)、（33,23)、（19,24)、（25,0)、（1，31)、（7,2)、·、n，1m、 (7,27)、（33，18)、（19，14)。值得注意的是，其他脈衝組^因^中 函數值太低而不列入次序串列，以降低搜尋時間。相較於 貝把例，弟一貫施例因脈衝組合的組合命中函數定義， 有不同的次序串列。 。^ 請參閱圖十四，圖十四為本發明之編碼向量的搜 程圖。本發明亦提供-種搜尋方法，肋於—語音編碼器伽^ encoder)中，搜尋一語音訊號(Speech signal)之一標的編碼向量 (Target codewector)。該語音訊號包含複數個編碼向量，該等編碼< 向量分別定義出複數個脈衝位置(pulse p0siti〇n)並且包令 喊 衝(Pulse)，每一脈衝係可分配在位於該編碼向量的該等脈衝 内，該等脈衝位置係分佈於複數個執道(Track)上。本發明之 方法包含下列步驟： ' 步驟S400 ·計算每一脈衝位置的一相對應命中函數 function)； 步驟S402 :根據每一執道中的脈衝位置個數與脈衝個數，決 定每一軌道中的複數組脈衝組合(Pulse combination); 17 1253057 步驟s·:根據每-脈衝位置的相對應命巾函數，計算每一 脈衝組合的-相對應組合命中函數(C()mbinati()n hit toetiQn); 步驟S406 :根據每-脈衝組合的相對應組合命中函數大小， 依序排列每-執這巾的脈衝組合，選擇每—執道巾具有最大組合 命中函數”，衝:且合二作為一相對應預選脈衝組合— p^e combmatK^，亚將其他脈衝組合依大小次序排列成一次序 ^驟S408 :根據每-執道中的預選脈衝組合，形成一預選編 ;〇de'VeCt〇r) 5 選脈ίϊ: 串列的順序’選擇次一脈衝組合為一候 合相對應^予暫時替代候選脈衝組 -e-vector)， 向量的決策分數f 決，分數，否低於預選編碼 S414 ； 乂驟S416，若否，則進行步驟 預選脈峨轨道的 步驟S416 ··檢查目前搜尋级 S410 ； 是，則進行步驟S418，若否合一預定搜尋條件，若 搜尋 步驟⑽：以目前的預選編碼向量為標的編碼向量，並完成 請參閱圖十五，圖十五為 實施例與AMR標準的代麵t 二第「實施例、第二 岭技哥方法的比較表。amr標 18 1253057 準搜尋1024次，第—每 — 次。由實驗結果得知? 二第二實施例各搜尋脈衝組合35 使用5.55秒編碼時間，本^^語音長度為12.32秒，AMR標準 時間，本發明之第二每/明之第—實施例使用4_57秒編碼 明之第-實施例f用05秒編碼時間，因此本發 22.7%的語音編碼執標準降低17.1〇/〇與 〇顧與〇.〇89，人 上品質PESQ值僅略為下降 衝組合取代技術，在兼顧任二分餅出差J。因此，本發明使用脈 雜度。 θ σσ貝的前提下，有效地降低計算複 少搜尋較佳^=^^據上述之實驗結果，本發明不僅可減« 差，進而確縮小原始語音訊號與編碼語音訊號的誤 發明具體實施例之詳述，係希望能更加清楚描述本 本發明之並f以上述所揭露的較佳具體實施例來對 及罝相Ϊ1 限制。相反地，其目的是希望能涵蓋各種改變 /、寺性的安排於本發明所欲申請之專利範圍的範疇内。Kn^p-bCn2). In this embodiment, the smaller the hit function value b(ni) or the combined hit function value, the earlier the order. As shown in Figure 12, the two (large, upper, and pulse combinations are sequentially (25, 25), (25, 30), (〇, 25), and (5 丁 0 pulse combinations (25, 25) The combined hit function value is 〇, pulse combination; ^2 where the hit function value is 1, and so on. Therefore, when the fourth device 38))== combination combination hit function, the pulses in each detour are arranged in order. The pulse combination of the first position hit function in the person's channel is the °^ of the orbital bit 0, the (1,1) of the mouth, the (7,7) of the τ2, and the τ3 of the trajectory (33, 33), and 执, 九, (19, 19). The above five pulse combinations are pre-selected pulse combinations, as shown in Figure 4, No. In this embodiment, other pulse combinations are sequentially ordered according to the position of the track = As shown in Figure 13 (Β), the order sequence is in order (25, 3 = sequence 丨 (7, 22), (33, 23), (19, 24), (25, 0), (1 , 31), (7, 2), ·, n, 1m, (7,27), (33,18), (19,14). It is worth noting that the value of the other pulse group ^ is too low It is not included in the order series to reduce the search time. Compared with the case of the case, the brother consistently applies the pulse combination. The combination hit function definition has a different order string. ^ Please refer to FIG. 14 and FIG. 14 is a search chart of the coding vector of the present invention. The present invention also provides a search method, which is a rib-to-speech encoder. ^ encoder), searching for a target codewector of a speech signal. The speech signal includes a plurality of coding vectors, and the codes respectively define a plurality of pulse positions (pulse p0siti〇n And the pulse is pulsed, each pulse can be allocated in the pulses located in the code vector, the pulse positions are distributed on a plurality of tracks. The method of the present invention comprises the following steps : ' Step S400 · Calculate a corresponding hit function for each pulse position); Step S402: Determine the complex array pulse combination in each track according to the number of pulse positions and the number of pulses in each track (Pulse combination 17 1253057 Step s:: Calculate the corresponding combination hit function (C() mbinati() n hit toetiQn) for each pulse combination according to the corresponding life function of each pulse position; Step S 406: according to the corresponding combination hit function size of each-pulse combination, sequentially arranging the pulse combination of each towel, selecting each towel to have the largest combined hit function, and rushing: and combining as a corresponding pre-selection Pulse combination - p^e combmatK^, sub-combining other pulse combinations into a sequence according to the order of magnitude S408: forming a pre-selection according to the pre-selected pulse combination in each-execution; 〇de'VeCt〇r) 5 ϊ 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 串 416 416 416 416 416 416 416 416 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 选择 416 416 416 416 416 416 416 416 416 If not, the step S416 of performing the pre-selection of the pulse track is to check the current search level S410; if yes, proceed to step S418, if the predetermined search condition is not combined, if the search step (10): the current pre-selected coding vector is used as the target Encoding vector, and refer to FIG. 15 for completion. FIG. 15 is a comparison table of the embodiment and the second embodiment of the AMR standard. Amr standard 18 1253057 Search for 1024 times, the first - every - times. It is known from the experimental results. In the second embodiment, each search pulse combination 35 uses 5.55 seconds of coding time, the length of the voice is 12.32 seconds, and the AMR standard time, the second embodiment of the present invention uses 4_57 seconds. In the coded example - the embodiment f uses the 05 second coding time, so the 22.7% speech coding standard of the present invention is reduced by 17.1 〇 / 〇 and care and 〇. 〇 89, the human quality PESQ value is only slightly reduced and combined with the technology. In the two-part biscuit business trip. Therefore, the present invention uses pulse complexity. Under the premise of θ σσ, effectively reducing the computational complexity of the search is better. According to the above experimental results, the present invention can not only reduce the difference, but also reduce the original embodiment of the original speech signal and the encoded speech signal. DETAILED DESCRIPTION OF THE INVENTION The present invention is intended to be more clearly described in the preferred embodiments of the present invention. On the contrary, it is intended to cover various changes and/or temples within the scope of the patent application to which the present invention is intended.
19 1253057 [圖式簡單說明】 圖〇 圖一為習知ACELP語音編碼器的示咅 圖一為根據AMR標準的】2 , 量的脈衝位置分佈圖。μ 12·21^模式中’一編碼向 圖 。圖三為習知技mm Am m之最狀鱗法之流程 圖四為習知技術之脈衝取代搜尋法之流程圖。 圖五為習知技術之次代數編碼薄搜尋法之流程圖。 之命狀搜尋法，於各軌道中，依各脈衝位置 P 〃數值大小排列的脈衝出現機率分佈圖。 Θ七為本务明之編碼向量的搜尋系統示意圖。 圖八為圖七中之第七裝置示意圖。 圖九為各崎位置的命巾函數示意圖。 示意圖 中函本發_第—實施例，所有脈衝組合及其組合命 =十（Α)為根據本發明的第一實施例所決定的預選脈衝組合 圖 。圖十（Β)為根據本發明的帛一實施例所決定的次序串列 不意 圖^一么ι 示意圖 合命中函本發明的第二實施例’所有脈衝組合及其組 20 1253057 圖十三(A)為根據本發明的第二實施例所決定的預選脈衝組合 示意圖。 圖十三(B)為根據本發明的第二實施例所決定的次序串列示意 圖。 圖十四為本發明之編碼向量的搜尋方法流程圖。 圖十五為比較本發明之第一實施例、第二實施例與AMR 標準的代數編碼薄搜尋方法的比較表。 【主要元件符號說明】 10 : ACELP語音編碼器 12 ： 前置處理器 14 :線性預估分析器 16 ： 適應性編碼薄搜尋器 18 :代數編碼薄搜尋器 20 ： 高頻濾波器 s(n):語音訊號 X⑻ :激發訊號 30 :搜尋系統 32 ： 第一裝置 34 :第二裝置 36 ： 第三裝置 38 :第四裝置 40 :第五裝置 42 :第六裝置 44 ： 第七裝置 46 :第一模組 48 ： 第二模組19 1253057 [Simple diagram of the diagram] Figure 1 is a diagram of a conventional ACELP speech coder. Figure 1 is a plot of the pulse position distribution according to the AMR standard. In the μ 12·21^ mode, a coded graph is shown. Fig. 3 is a flow chart of the most accurate scale method of the conventional technique mm Am m. Fig. 4 is a flow chart of the pulse substitution search method of the prior art. FIG. 5 is a flow chart of a sub-algebraic codebook search method of the prior art. According to the life search method, in each track, the probability distribution pattern of the pulses arranged according to the value of each pulse position P 〃 is shown. Θ7 is a schematic diagram of the search system for the code vector. Figure 8 is a schematic diagram of the seventh device in Figure 7. Figure 9 is a schematic diagram of the life towel function at each position. In the schematic diagram, the letter-first embodiment, all combinations of pulses and their combinations = ten (Α) are the preselected pulse combination maps determined in accordance with the first embodiment of the present invention. Figure 10 (Β) is a sequence of sequences determined according to an embodiment of the present invention. It is not intended to be a schematic diagram of a second embodiment of the invention. All of the pulse combinations and their groups 20 1253057 Figure 13 ( A) is a schematic diagram of a preselected pulse combination determined in accordance with a second embodiment of the present invention. Figure 13 (B) is a schematic diagram of a sequence of sequences determined in accordance with a second embodiment of the present invention. FIG. 14 is a flow chart of a method for searching for a code vector according to the present invention. Figure 15 is a comparison table comparing the algebraic codebook search methods of the first embodiment, the second embodiment and the AMR standard of the present invention. [Main component symbol description] 10 : ACELP speech encoder 12 : Pre-processor 14 : linear prediction analyzer 16 : Adaptive code thin searcher 18 : Algebraic code thin searcher 20 : High-frequency filter s (n) : voice signal X (8): excitation signal 30: search system 32: first device 34: second device 36: third device 38: fourth device 40: fifth device 42: sixth device 44: seventh device 46: first Module 48: second module