200917230 九、發明說明： 【發明所屬之技術領域】 本發明係提供一種用於語音編碼器增益碼薄分段搜尋演 算法’此尋演算法可應用於語音通訊產業，用以減少編碼器計 算量之演算法。 【先前技術】 按，目前世界多個標準組織和工業實體提出了很多話音編 碼方案。其中’包括有國際電信聯盟（ITU-T)的G.723.1編 碼方案。而現有文獻的技術，由&ing-Kyo Jung等人所發表之 A fast adaptive-codebook search algorithm for G.723.1 speech coder”（用於G. 723.1語音編碼器之快速適應性碼薄 氣哥展务法、，IEEE Signal processing letters, ΊΙ. VI、 N〇. 1，pp. 75-78，January 2005。其係將傳統的 g. 723.1 五200917230 IX. Description of the Invention: [Technical Field] The present invention provides a speech encoder gain code segmentation search algorithm 'This search algorithm can be applied to the voice communication industry to reduce the amount of encoder calculation Algorithm. [Prior Art] According to the current, many standard organizations and industrial entities in the world have proposed a number of voice coding schemes. Among them, the G.723.1 coding scheme of the International Telecommunication Union (ITU-T) is included. The technique of the prior art, A fast adaptive-codebook search algorithm for G.723.1 speech coder" by &ing-Kyo Jung et al. (for the fast adaptability of the G. 723.1 speech coder) Law, IEEE Signal processing letters, ΊΙ. VI, N〇. 1, pp. 75-78, January 2005. Its department will be traditional g. 723.1 five
階閉迴路適應性基週的增益碼薄做平均分段範圍來搜尋，雖然 其计异複雜度較傳統G. 723.1全部搜尋約改善5〇%，但此語音 編碼器技術之計算複雜度與語音品質無法達到最佳調配。而 且’傳統G. 723.1五階閉迴路適應性碼薄基週的增益之搜尋是 採用碼薄全部搜尋，這種技術之計算複雜度最高，以致於CPU 負荷最大，不適合低階廉價的CPU之手機、舰、掌上型電腦 等裝置產品。 關於語音編碼器之技術’本發明人已有申請核准台灣發明 專利公告第5_4號「聲音聽覺雜使用於殘鋪換可變速 200917230 率語音編解碼器」，在此併人本文，以供參考。 【發明内容】 明之目的係提供—翻於語音編彻增益碼薄分段 ，寅开法可用以減少編碼器計算量，進而可選用較低階的 (PU運算核心，達猶低產品關發量產成本、及因改善編媽 器需求的計算量達到降低電池耗電量。 α為達致上述目的，本發明用於語音編碼器增益碼薄分段搜 尋演算法，包括：提供一碼薄（五階閉迴路適應性碼薄），包 含有複數個碼向量㈤eve伽s)，每—碼向量包含有複數個 要素（elements)，·將該前述之碼薄分為前後兩段，前半段為 一前碼簿（Pre-codebook )，後半段為一後碼薄 (Post-codebook)，並由一門檻值（Thresh〇ld)以 a 值代表 作為決定所分段之碼薄大小，該前碼薄〇 一a)乘上一邊界值（B0und)，而該後碼薄（P〇st—c〇deb〇〇k) 為a乘上該邊界值，其中，該邊界值（B〇und)為前述該碼薄 之大小；而該a值介於大於0.5小於1之範圍内。更可包括事 先（predetermined)利用一階的閉迴路基週來預估該五階閉 迴路適應性碼簿基週的增益之搜尋範圍，藉以來達到降低搜尋 演算法計算量。 為了讓本發明之上述目的、特徵、優點能更明顯，下文特 舉本發明較佳實施例，並配合所附圖示，作詳細說明如下。 【實施方式】 200917230 請參閱第一圖’其係本發明實施例五階閉迴路適應性基週 的增益碼薄做不均勻分段範圍之示意圖。包括有： 一碼薄ίο，例如為一適應性增益碼薄（gaincodebook)， 其包含有衩數個碼向量12 (codevectors )，於本實施例係以 170個碼向量碼薄來說明，由k=〇，k=l,......至k=169來表示， 且係以一階增從碼向量（1 - tap ga i n c〇devect〇r)為例說明 (如第一圖中箭頭所示之位置）。每一碼向量更包含有複數個 要素14 (elements)，於本實施例係以20個要素說明，由 g〇, gl，g2......至 gl9 來表示。 將該前述之碼薄10分為前後兩段，前半段為一前碼薄14 (Pre-codebook) ’ 後半段為一後碼薄 16 (Post-codebook)， 並由一門檻值（Threshold)以（a)值代表作為決定所分段之 碼溥大小。該前碼薄（Pre-codebook)之大小定義如下： Pre-codebook = (1 —a) X Bound。 後碼薄（Post-codebook)之大小則為 Post-codebook = a x Bound ° 其中，邊界值（Bound)為原適應性增益碼薄之大小。 本發明實施例係利用一標準適應性增益碼薄（Standard Adaptive Codebook search，簡稱Standard ACB search)搜 尋最佳參數之判斷準則，以目標訊號5^1與預估訊號#"】所計 算之最小均方誤差（Minimum Square Error, MSE)來評估適應 200917230 性碼薄預估週期性激發訊號之準確性，MSE越低，則表示適應 性碼薄搜尋的預估訊號越準確。 以下所顯示之表一，其係以三段男生（Male#1，Male#2, Male #3)及三段女生（Female Female #2，Fe贴化 之浯音汛號（Speech signal)為實驗測試訊號，並提出平均 值（Average) ’針對不同的Threshold評:丨古其適應性增益碼向 量預估的準確性分析，a = 〇· 5即是現有習知的文獻技術將適 應性增益碼薄平均分段搜尋。本發明經實驗分析的結果顯示將 碼溥作不均勻分段搜尋且a值越大，其適應性增益褐向量搜尋 預估的準確度是較平均分段為佳的。為了使編碼器計算複雜度 與合成語音品質之間達到最佳調配，所提出的增益碼薄不均勻 分段快速搜尋法，在本實施例係選定以a = 〇· 7將增益碼薄不 均勻;7 #又，可達到語音編碼器較有效率編碼。經表—之數據顯 不’ a >〇. 5時所得之分段區域皆優於a =〇· 5之平均分段。 不同的Tlireshold(〇!)設定值其MSE之分析 Speech — Gain codebook tkeshold (a) signal I 01 I 0.2 | 0.3 | 0.4 丨05 1 0.6 0.7 厂0.8 1 0.9 Female #1 5.002 5.016 5.002 4.9S4 4.993 4.981 4.983 4.984 4.971 Female #2 4.404 4.387 4.396 4.393 4.394 4.380 4.368 4.383 4.375 Female #3 5.842 5.831 5.S21 5.829 5.818 5.816 5.806 5.S14 5.809 Λ-iale #1 5.101 5.065 5.083 5.077 5.062 5.069 5.070 5.067 5.062 Male #2 5.750 5.740 5.740 5.737 5.741 5.734 5.732 5.735 5.720 Male #3 5.491 7.483 1.484 rs.m 5.479 5.470 5.47J 5.459 5.461 Average 5,265 | 5.254 .... 5.251 5·纖 „5.242 5,238 5.240 5 233 200917230 以下所顯示之表二，其係本發明實施例之語音品質訊號雜 訊比（Signal Noise Ratio,簡稱 SNR)評量。以原 G. 723.1 標準適應性碼薄搜尋法（Standard ACB search)、現有習知的 文獻技術：快速適應性碼薄搜尋法（FAST ACB search， Threshold = 0.5)及本發明所提出的增益碼薄不均勻分段搜 寻法（Proposed method，Threshold = 0. 7)。等三種搜尋演 鼻法之效能s平估欄位（performance 在效能評估攔位（performance evaluations)上包含有 兩大主項，分別為SNR (db)評估，是作為客觀判定語音品質 好壞的比值’以分貝數（db)為單位來表示，若SNR值越高， 則表示語音品質越好’反之則越差；以及SNR降低 (degradation)之百分比（％)’上述之判定對象包含有男生 (Male)、女生（Female)及其平均值（Average)。由表二得 知本發明所提出的適應性增益碼薄不均勻分段快速搜尋法比 習知文獻所提出將適應性增益碼薄平均分段快速搜尋法有更 佳的合成語音品質。與G. 723.1標準適應性碼薄搜尋演算法相 較之下，本方法其平均SNR值約只降低了 〇. 〇94 dB。 其中’表二所示之 Standard ACB search :原 G. 723.1 標 準適應性瑪簿搜尋演算法。 文獻技術：Fast ACB search (Threshold = 0. 5)與本發 200917230 明所提出的增益碼薄不均勻分段搜尋法其合成語音品質SNR 之評估。 表二The closed-loop adaptive base-cycle gain codebook is used to search for the average segmentation range. Although its calculation complexity is about 5% higher than the traditional G.723.1 search, the computational complexity and speech of this speech coder technology Quality cannot be optimally matched. Moreover, the search for the gain of the traditional G. 723.1 fifth-order closed-loop adaptive codebook base is based on the full search of the codebook. This technique has the highest computational complexity, so that the CPU load is the largest, and it is not suitable for the mobile phone with low-order and low-cost CPU. , ship, palm-sized computers and other device products. Regarding the technology of the speech coder, the inventors have applied for approval of the Taiwan invention patent publication No. 5_4 "Sound and Hearing Miscellaneous for Remnant and Variable Shifting 200917230 Rate Speech Codec", which is hereby incorporated by reference. SUMMARY OF THE INVENTION The purpose of the present invention is to provide a method for turning over the speech coded gain code thin segment, and the cleaving method can be used to reduce the amount of calculation of the encoder, and then a lower order (PU operation core can be used to achieve a low product shutdown amount). The production cost and the calculation amount of the demand for improving the device are reduced to reduce the battery power consumption. α In order to achieve the above object, the present invention is applied to a speech encoder gain code segmentation search algorithm, including: providing a codebook ( The fifth-order closed-loop adaptive codebook) includes a plurality of code vectors (five) eve gamma s), each code vector includes a plurality of elements, and the foregoing codebook is divided into two segments, the first half is A pre-codebook, the latter half is a post-codebook, and is represented by a threshold (Thresh〇ld) with a value representing the size of the code segment segmented as the decision. The thin layer a) is multiplied by a boundary value (B0und), and the latter codebook (P〇st-c〇deb〇〇k) multiplies a by the boundary value, wherein the boundary value (B〇und) is The size of the aforementioned codebook; and the value of a is greater than 0.5 and less than 1. It may further include predetermining the first-order closed-loop base period to estimate the search range of the gain of the fifth-order closed-loop adaptive codebook base period, thereby reducing the calculation amount of the search algorithm. The above described objects, features and advantages of the present invention will become more apparent from the detailed description of the preferred embodiments of the invention. [Embodiment] 200917230 Please refer to the first figure', which is a schematic diagram of a non-uniform segmentation range of a gain codebook of a fifth-order closed-loop adaptive base period according to an embodiment of the present invention. Including: a codebook ίο, for example, an adaptive gain codebook (gaincodebook), which includes a plurality of code vectors (codevectors), which is illustrated by 170 code vector codebooks in this embodiment, by k =〇,k=l,... to k=169 to represent, and the first-order incremental code vector (1 - tap ga inc〇devect〇r) is used as an example (such as the arrow in the first figure) The location shown). Each code vector further includes a plurality of elements 14 (elements), which are described by 20 elements in this embodiment, and are represented by g〇, gl, g2, ... to gl9. The aforementioned codebook 10 is divided into two segments, the first half is a pre-codebook 14 and the second half is a post-codebook, and is thresholded by a threshold. (a) The value represents the size of the pallet as the decision. The size of the pre-codebook is defined as follows: Pre-codebook = (1 - a) X Bound. The size of the post-codebook is Post-codebook = a x Bound ° where the boundary value (Bound) is the size of the original adaptive gain codebook. The embodiment of the present invention uses a standard adaptive codebook search (Standard ACB search) to search for the optimal parameter, and the minimum value calculated by the target signal 5^1 and the estimated signal #" The Minimum Square Error (MSE) is used to evaluate the accuracy of the periodic excitation signal for the 200917230 codebook. The lower the MSE, the more accurate the prediction signal for the adaptive codebook search. Table 1 shown below is based on three male students (Male #1, Male #2, Male #3) and three female students (Female Female #2, Fe Sticking's Speech Signal). Test the signal and propose the average (Average) 'For different Threshold reviews: The accuracy analysis of the adaptive gain code vector estimation of the 丨古古, a = 〇 · 5 is the existing knowledge of the literature technology will adapt the gain code Thin average segmentation search. The experimental analysis results show that the code is used for uneven segmentation search and the larger the a value, the accuracy of the adaptive gain brown vector search prediction is better than the average segmentation. In order to achieve optimal matching between the computational complexity of the encoder and the quality of the synthesized speech, the proposed gain code thin uneven segmentation fast search method is selected in this embodiment to have a non-uniform gain code thinness with a = 〇·7. ; 7 # again, can achieve a more efficient coding of the speech encoder. The data obtained by the table - not a ' a > 〇. 5 segmentation area is better than the average segment of a = 〇 · 5. Different Tlireshold (〇!) set value of its MSE analysis Speech — Gain codebook Tkeshold (a) signal I 01 I 0.2 | 0.3 | 0.4 丨05 1 0.6 0.7 Plant 0.8 1 0.9 Female #1 5.002 5.016 5.002 4.9S4 4.993 4.981 4.983 4.984 4.971 Female #2 4.404 4.387 4.396 4.393 4.394 4.380 4.368 4.383 4.375 Female #3 5.842 5.831 5.S21 5.829 5.818 5.816 5.806 5.S14 5.809 Λ-iale #1 5.101 5.065 5.083 5.077 5.062 5.069 5.070 5.067 5.062 Male #2 5.750 5.740 5.740 5.737 5.741 5.734 5.732 5.735 5.720 Male #3 5.491 7.483 1.484 rs.m 5.479 5.470 5.47J 5.459 5.461 Average 5,265 | 5.254 .... 5.251 5·fiber „5.242 5,238 5.240 5 233 200917230 Table 2, which is shown below, is a speech quality noise ratio (Signal Noise Ratio, SNR for short) ) Evaluation. The original G. 723.1 standard adaptive code search method (Standard ACB search), the existing literature technology: fast adaptive code search method (FAST ACB search, Threshold = 0.5) and the proposed gain codebook Proposed method (Threshold = 0.7). The performance of the three search for the performance of the nasal smearing field (performance in performance evaluations (performance evaluations) contains two main items, respectively, SNR (db) evaluation, as an objective judgment of the quality of speech quality ratio 'Denoted in decibels (db), if the higher the SNR value, the better the speech quality is. 'The worse the worse; and the percentage of SNR degradation (%)' The above judgment object includes boys ( Male), Female (Africa) and its average (Average). It is known from Table 2 that the adaptive gain codebook uneven segmentation fast search method proposed by the present invention proposes an adaptive gain codebook average than the conventional literature. The segmented fast search method has better synthesized speech quality. Compared with the G. 723.1 standard adaptive codebook search algorithm, the average SNR value of this method is only reduced by 〇. 〇94 dB. Standard ACB search: original G. 723.1 standard adaptive marathon search algorithm. Literature technology: Fast ACB search (Threshold = 0. 5) and the present invention 200917230 Ming proposed gain code thin uneven segmentation search Assess the quality SNR of the synthesized speech. Table II
以下所顯示之表三，其係本發明實施例為不同適應性碼薄 搜尋演算法所需加法（Additions)乘法（MultipHcations) 數（170個碼向量）之比較表。不同演算法欄位（DifferentTable 3, which is shown below, is a comparison table of the number of additions (Multiplications) required by the different adaptive codebook search algorithms (170 code vectors). Different algorithm fields (Different
Algorithm)包括有：G· 723. 1標準適應性碼簿搜尋法(货andard ACB search)、現有習知的文獻技術，快速適應性碼薄搜尋法 (FAST ACB search)，其係採用均勻分段（Eve請) 及本發明所提出的增益瑪薄不均勻分段搜尋法（uneven segmentation ACB search)，上述三種碼薄搜尋演算法是採用 170個碼向量之碼薄，比較這三種碼薄搜尋演算法所需要的加 法、乘法數。所提出的改良方法是以α = 〇 7將增益碼薄做 不均勻分段搜尋。以本發明最差（worst case)的情況而言， 換έ之也就是一階濾波器初步預估之增益碼向量g都落在 Post-codebook之内，其搜尋範圍為整個適應性增益碼薄之 200917230 70%，而所減少的計算量仍高達30%。考慮所加入的一階基週 增益初步預估，對整體適應性碼薄的搜尋而言，仍然可以節省 20%以上的計算量。 表三 不同適應性碼薄搜尋演算法所f加法乘法數(170個碼向量)Algorithm) includes: G·723. 1 standard adaptive codebook search method (goods andards ACB search), existing document technology, fast adaptive code search method (FAST ACB search), which uses uniform segmentation (Eve please) and the uneven segmentation search method (uneven segmentation ACB search) proposed by the present invention, the above three codebook search algorithms use 170 code vector codebooks to compare the three codebook search calculus The number of additions and multiplications required by the law. The proposed improvement method is to perform uneven segmentation search on the gain codebook with α = 〇 7. In the case of the worst case of the present invention, the gain code vector g of the first-order filter is estimated to fall within the Post-codebook, and the search range is the entire adaptive gain codebook. The 200917230 is 70%, and the amount of calculations reduced is still as high as 30%. Considering the initial estimate of the first-order base-cycle gain added, it can still save more than 20% of the calculation for the overall adaptive codebook search. Table 3 Addition multiplication numbers (170 code vectors) for different adaptive codebook search algorithms
Different Algoriilun Additions Multiplications Standard ACB search 311.10K 953.70K Fast ACB search(Even segmentation) 186.66K 572.22K Uneven SCiUtien h on a = 〇 7 Best case 124.44K 381.48K ACB search Worst case 248.88K 762.96K 如剷所述’本發明已提出一個新的快速搜尋法來進一步的 改善減>、傳統的G. 723.1語音編瑪器適應性增益碼薄的搜尋 演算法計算娜度。本發明之演算法是贿麟五階適應性增 益碼薄做不均分段且事先（predeterfflined)糊—階的閉^ 路基週來預估傳統G.723」五_迴路適應性碼薄基週的增 益之搜尋範圍’藉絲達浙健尋演算法計算量，所以本發 明之快速搜尋演算法之計算额語達到紐 調配，藉以提高搜尋最佳增益的料度。客觀的評估編碼器之 效能，於適應性增益向量碼薄的SNR值約只降低了 〇._仙， 然而卻可以節省適應性增益碼薄之搜尋演算法計算量達挪以 200917230 上0 1.可應用在手持式行動裝置(如手機、m、掌 等語音通财面，可㈣齡編碼_算量触 的嘱算核心達到降低產品的開發量產成本、及因改善= 器需求的計算#達到降低電池耗電量。 2.本發明之技術可應用在醬網路電話系統上，減少語 音編碼器的計算量。 ° 故本發明之提出，應符合專利產業上利用性、新穎性、以 及進步性之所規定。軸前賴描述及圖式已揭示本發明之較 佳實施例，惟此乃僅係實細之呈現，舉凡各種增添、修改和 取代可能使用於本發明較佳實施例，仍應屬落入本發明之申請 專利範圍所界定之範圍内。因此，本文於此所揭示的實施例所 有觀點，應被視為用以說明本發明，而非用以限制本發明。本 發明之範圍應由後附之申請專利範圍所界定，並涵蓋其合法均 等物，並不限於先前之描述。 11 200917230 【圖式簡單說明】 第一圖係本發明實施例五階閉迴路適應性基週的增益碼薄做 不均勻分段範圍之示意圖。 【主要元件符號說明】 10 碼薄 12 碼向量 14 要素 16 前碼薄 18 後碼薄 12Different Algoriilun Additions Multiplications Standard ACB search 311.10K 953.70K Fast ACB search(Even segmentation) 186.66K 572.22K Uneven SCiUtien h on a = 〇7 Best case 124.44K 381.48K ACB search Worst case 248.88K 762.96K The invention has proposed a new fast search method to further improve the subtraction, and the traditional G. 723.1 speech coder adaptive gain codebook search algorithm calculates the degree. The algorithm of the present invention is to estimate the traditional G.723" five-loop adaptive code base period by using the fifth-order adaptive gain codebook to be unevenly segmented and predeterfflined. The search range of the gain is calculated by the calculation method of the sifting algorithm, so the calculation of the fast search algorithm of the present invention reaches the Newton allocation, thereby improving the search for the optimal gain. Objectively evaluate the performance of the encoder. The SNR value of the adaptive gain vector codebook is only reduced by 〇._仙, but the adaptive gain codebook search algorithm can be saved. The calculation amount is up to 200917230. It can be applied to hand-held mobile devices (such as mobile phones, m, palm, etc.), which can be used to reduce the development and mass production cost of products and the calculation of the demand for improved devices. The battery power consumption can be reduced. 2. The technology of the present invention can be applied to the sauce network telephone system to reduce the calculation amount of the voice encoder. Therefore, the present invention is proposed to be compatible with the patent industry, novelty, and The present invention has been described in terms of a preferred embodiment of the present invention, and is intended to be a It is intended to be within the scope of the invention as defined by the appended claims. The scope of the invention should be defined by the scope of the patent application, and its legal equivalents are not limited to the previous description. 11 200917230 [Simplified description of the drawings] The first figure is the fifth-order closed-loop adaptive basis of the embodiment of the present invention. Schematic diagram of the week's gain code thinning to make the uneven segmentation range. [Main component symbol description] 10 code thin 12 code vector 14 element 16 front code thin 18 post code thin 12