2 的 438 五、發明説明（1 ) 發明範固 本發明係有關於一船站鉦― 股地聲音辨識及高度的或易變 環境中特殊地聲音辨識。 的噪 發明背景 進步的技*正藉由像語音辨識器等電子裝 商業化聲音辨識邁進°大致有兩種型式之語音辨識器。^ 種會在使用讀人短的命令時執行某種㈣1二種 接受口述語音並如文字般的輸入語音。 7 訂 複數的語音辨識器在它們能辨識使用者所説單字 前必須先受使用者訓練。這些被稱爲"特定語者"的語音: 識器，意思是在語音辨識器能解析使用者單字和命^芦 音辨識器必須接受使用者發聲訓練。要訓練語音辨^器: 常需要使用者多次的哈某些單字或片語輸入辨識器中，如 此語音辨識器可以辨識出使用者説話的模型。當使用者^ 用語音辨識器時，語音辨識器將會比對輸入聲音信號與不 同的浯T樣板以找到最類似輸入聲音信號之樣板。這種方 法被稱爲"模型匹配·，。 經濟部中央標準局員工消費合作社印— 使用者通常將在相對地低干擾噪音環境中，，訓練"語音辨 識器。後來，複數的語音辨識器就一定要使用在低干擾噪 音環境中。不然，語音辨識器將無法自背景噪音中分辨出 所説的單字。若語音辨識器使用在低噪音環境中，將可達 成適當高的辨識率。如果語音辨識器在具有一適度的、不 變的背景噪音的位置接受訓練，且後來用於具有同樣適度2 of 438 V. Description of the invention (1) The invention of the invention The invention relates to the sound recognition of a ship's station-the sound recognition and the special sound recognition in a highly or variable environment. The background of the invention is progressing through electronic devices such as speech recognizers. Commercial voice recognition has advanced. There are roughly two types of speech recognizers. ^ One will perform a certain type of (2) when using short commands to read people. It accepts spoken speech and enters the speech like text. 7 The plural speech recognizers must be trained by the user before they can recognize the words spoken by the user. These voices called "quote-specific speakers": recognizers, which means that the speech recognizer can parse the user's words and commands ^ Lu recognizer must receive user vocal training. To train a speech recognizer: It often requires the user to input certain words or phrases into the recognizer many times, so that the speech recognizer can recognize the model of the user's speech. When the user ^ uses a speech recognizer, the speech recognizer will compare the input sound signal with different reticle templates to find the most similar input sound signal template. This method is called " model matching. Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economy-Users will usually train " speech recognizers in relatively low-noise environments. Later, plural speech recognizers must be used in a low-interference noise environment. Otherwise, the speech recognizer will not be able to distinguish the spoken word from the background noise. If the speech recognizer is used in a low-noise environment, it will achieve a suitably high recognition rate. If the speech recognizer is trained in a position with a moderate, unchanging background noise, and later used to have the same moderate
A7 經 濟 部 中 標 準 局 員 工 消 費 合 作 社 印 製 五、發明説明（ 的、不變的背哥唾立 而，當這些語音辨日、％’將可達成高的辨識率。然 日辨硪器使用在具有負信號噪古 =境及與出現在訓練期間背景嗓音不同噪音d 識率會降得非常低’沒有用的準確水準。要修：:::噪： 的問題，傳統的語音辨識器嘗試預估: ::定使用者聲音上的影響。不同的技術被::: = 中減去之噪音的統計或參數 變的環境中，這些模型是非常不正確的。“度的、易 附圖之簡單説明 圖1是根據本發明最#具體實例的聲音辨識器之方 圖。 圖2表示本發明最佳具體實例的流程圖。 圖3表不根據本發明最佳具體實例使用計算特徵値 之流程圖。 圖4是根據本發明利用濾頻器處理的取樣 譜描述圖。 B ° 圖5表示根據本發明最佳具體實 徵矩陣。 •I γ '装！ (請先閱讀背面之注意事項再填寫本頁) 塊 方法 號之能 例的取樣聲音信號之特 圖6表示根據本發明最佳具體實例圖3之特徵的正規化特 徵矩陣。 發明之詳細説明 本發明最佳具體實例被使用於聲音辨識器之健全語立 辨 -丁 、τ ( cns ) (210X297公釐 經濟部中央標準局員工消費合作社印製 五、發明説明（3 ) 識。最佳具體實例很適合使用於汽車上的蜂巢式電話 '吏用者雙手保持在方向盤上、雙眼看著路且仍能在窗户開 著且大聲放著音響下打電話。 1 1豕杜间且/或易變的噪音 條件下準確率不良的傳統語音辨識器，根據本發明最佳具 體實例所設計聲音辨識器是健全的且在易變噪音和梁音程 度大於使用者説話音量的環境中得到非常高的準確率。 本發明將錢文結合附圖詳加描述。特別地，最佳具體 實例將參考圖1並結合其他圖示加以描述。 本發明可以應用至任何音響聲音的辨識。例如，此音響 聲音可能是語音、咕嚼聲、動物所發出的聲音、包括打擊 樂器等樂器的聲音或其他形式聲音。大體來説，本發明係 有關語音的辨識。 圖1表不根據本發明最佳具體實例的聲音辨識器1 。最 佳具體實例中’一音響信號輸入至聲音辨識器1〇〇的類比 數位轉換（ADC)在此信號轉換成數位信號且以丨6千赫的速 率取樣。其他適當的取樣速率也可能使用，諸如8千赫。 取樣數位信號輸入至將取樣數位信號分配給分析框之特 徵向量裝置110。各分析框可選擇爲固定時間寬度（如2〇微 秒）或可能是依據信號特性諸如音調周期或其他決定因子 之變動時間寬度。各分析框的起始點可選擇在前一框結束 點之前或之後。最佳具體實例中，分析框選擇爲固定時間 寬度’且各分析框的起始點起始於前一分析框之結束點。 由於各分析框，特徵向量裝置110計算特徵向量（圖2流 程圖之210)。任何指定數量之分析框，特徵向量裝置j 1() (請先閱讀背面之注意事項再填寫本頁) -策 —IT---------------- • I I I ·A7 Printed by the Consumer Standardization Bureau of the China Bureau of Standards, Ministry of Economic Affairs V. Description of the invention (The unchanging back brother stands, when these voice recognition days,% 'will achieve a high recognition rate. However, the daily recognition device is used in It has a negative signal noise environment and different noise from the background voice during training. The recognition rate will be very low 'useless accurate level. To solve the problem of :::: Noise: traditional speech recognizers try to predict Estimate: :: Determine the impact on the user's voice. Different technologies are ::: = in the environment where the noise statistics or parameters change, these models are very incorrect. "Degree, easy to attach Brief Description FIG. 1 is a block diagram of a voice recognizer according to the most specific example of the present invention. FIG. 2 shows a flowchart of the best specific example of the present invention. FIG. 3 shows a flow of using the calculation feature value according to the best specific example of the present invention Figure. Figure 4 is a description of the sampling spectrum processed by the frequency filter according to the present invention. B ° Figure 5 shows the best specific actual matrix according to the present invention. • I γ 'install! (Please read the precautions on the back before filling in This page) block The specific example of the sampled sound signal of the method number is shown in Fig. 6 as the normalized feature matrix of the features of Fig. 3 according to the best embodiment of the present invention. Detailed description of the invention The best embodiment of the present invention is used in the sound language of the sound recognizer Discrimination-Ding, τ (cns) (210X297mm, printed by the employee consumer cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 5. Invention description (3). The best specific examples are very suitable for the use of cellular telephones ’users in automobiles Keep your hands on the steering wheel, look at the road with your eyes and still be able to make calls while the windows are open and loudly on the speakers. 1 1 Traditional voice recognizer with poor accuracy under doudu and / or variable noise conditions The sound recognizer designed according to the best specific example of the present invention is sound and obtains a very high accuracy in an environment with variable noise and beam sounds greater than the user's speaking volume. The present invention adds Qian Wen in conjunction with the drawings Description. In particular, the best specific example will be described with reference to FIG. 1 in conjunction with other illustrations. The present invention can be applied to the recognition of any acoustic sound. For example, this acoustic sound may Speech, chewing sounds, sounds made by animals, sounds of instruments including percussion instruments and other forms of sound. Generally speaking, the present invention is related to the recognition of speech. FIG. 1 shows the sound recognition according to the best specific example of the present invention 1. In the best specific example, an analog digital conversion (ADC) where an audio signal is input to the sound recognizer 100 is converted into a digital signal and is sampled at a rate of 6 kHz. Other suitable sampling rates are also It may be used, such as 8 kHz. The sampled digital signal is input to the feature vector device 110 that distributes the sampled digital signal to the analysis frame. Each analysis frame may be selected to have a fixed time width (eg, 20 microseconds) or may be based on signal characteristics such The duration of the change in pitch period or other determinants. The starting point of each analysis frame can be selected before or after the end point of the previous frame. In the best specific example, the analysis frame is selected to have a fixed time width and the starting point of each analysis frame starts at the end point of the previous analysis frame. Due to each analysis frame, the feature vector device 110 calculates a feature vector (210 in the flowchart of FIG. 2). Any specified number of analysis frames, feature vector device j 1 () (please read the notes on the back before filling this page)-策 —IT ---------------- • I I I ·
.I I I -6- 本紙張尺度適用中國國豕標準（CNS ) Α4規格（210Χ297公釐） B7 五、發明説明（4 ) 產生等數量之特徵向量。一特徵向量是—自所指定分析框 内的取樣聲音信號推導之連續數値或許多的特徵値。這些 特徵値代表取樣聲音信號所包含的資料。 一 有許多精通此語音辨識技藝者所知可能被用來決定特徵 向量之技術。此技術包括線性預測編碼（Lpc)係數' CepstraH系數、對數區域比率及麥耳音階濾波器串係數。本 發明最佳具體實例使用麥耳音階濾波器_係數法，雖然本 發明將運用其他特徵向量技術，諸如上文所列的方法。 參考圖3流程圖，麥耳音階濾波器串係數法以下列方法 計算。 1.分析框之聲音信號取樣經由高頻前置加權濾波器白化 聲音信號取樣之光譜（圖3流程圖之31〇)。這樣增加了高頻 成分之相對能量與低頻成分之能量比對。當本發明最佳具 體實例使用語音信號可得到利益因爲語音低頻成分有一比 局頻成分大之相對能量，且此兩成分於前置加重濾波器中 再平衡。最佳具體實例中，根據以下方程式可完成濾波： 經濟部中央標準局員工消費合作社印製 (請先閲讀背面之注意事項再填寫本頁).I I I -6- This paper scale is applicable to China National Standards (CNS) Α4 specification (210Χ297mm) B7 5. Description of the invention (4) Generate equal number of feature vectors. A feature vector is a continuous number value or many feature values derived from the sampled sound signal in the specified analysis frame. These characteristic values represent the data contained in the sampled sound signal. There are many techniques known to those skilled in this speech recognition technique that may be used to determine feature vectors. This technique includes linear predictive coding (Lpc) coefficients' CepstraH coefficients, log domain ratios, and Maier scale filter string coefficients. The best embodiment of the invention uses the Maier scale filter_coefficient method, although the invention will use other feature vector techniques, such as the methods listed above. Referring to the flowchart of FIG. 3, the Maier scale filter string coefficient method is calculated in the following manner. 1. The sound signal sample of the analysis frame is whitened by the high-frequency pre-weighting filter. The spectrum of the sound signal sample (31 in the flowchart of FIG. 3). This increases the relative energy of the high-frequency component and the energy of the low-frequency component. When the best specific example of the present invention uses speech signals, benefits are obtained because the low frequency components of speech have a relative energy greater than the local frequency components, and these two components are rebalanced in the pre-emphasis filter. In the best specific example, the filtering can be completed according to the following equation: Printed by the Employees Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs (please read the precautions on the back before filling this page)
Pi(k) = s^k) - Si(k-l) 其中Si(k)是分析框"i"在位置k之聲音信號取樣，是分 析框"i"在前面位置時間"k·〗"之聲音信號取樣，且仍⑻是 分析框"i"在位置k之前置加重聲音信號取樣。一精於此聲 音辨識技術領域者將辨識其他前置加重濾波器可能被使 用。 本紙張尺度適用中國國家標隼（CNS )八4規格（21〇χ297公釐j 經濟部中央標準局員工消費合作社印製 A7 '_______B7 五、發明説明（5 ) ~ ~ ~~~ 2.各分析框之前置加重聲音信號取樣被—聯串轉換不同 頻帶之濾波器帶通濾波。此濾波器可應用於任何想要在時 域或頻域的計算方法。最佳具體實例中，攄波器應用於頻 域中。然而，首先，分析框中前置加重聲音信號取樣之能 譜必須計算（圖3之320)。能譜由下求得： a. 分析框中前置加重聲音信號取樣乘上觸發脈衝函數 或加權函數之取樣。任可觸發脈衝函數皆可能應用。爲了 解釋本發明，d可假設一簡單的矩形觸發脈衝（所有取樣之 觸發脈衝値爲1.0)。 b. 計算各觸發脈衝分析框中前置加重聲音信號取樣之 傅利葉轉換。 c. 傅利葉轉換値平方得到能譜値。 把譜値決疋後，帶通濾波器被各能譜値之濾波器加權値 應用於頻域（圖3之330)。雖然許多濾波器加權函數可能使 用於帶通濾波器中，最佳具體實例合併圖4中所見之上升 餘弦加權斷面。 圖4表示有利用上升餘弦斷面41〇在上面之能譜4〇〇。本發 明最佳具體實例中，根據接近人類耳朵頻率響應之麥耳或 巴克音階’各帶通濾波器或上升餘弦斷面41〇之頻帶沿著 頻率軸繪製。帶通濾波器之頻帶（上升餘弦斷面41〇)趨近 線性地由0至1千赫間隔，且l〇grithmically間隔1千赫以上。 那些最佳具體實例所定義之外的遽波間隔亦可使用。如圖 4之最佳具體實例所見’帶通濾波器或上升餘弦斷面41〇重 疊。帶通濾波器或上升餘弦斷面410輸出的計算是根據： 本紙張尺度適用中國國家標準（CNS ) M規格（210x297公釐） (請先閱讀背面之注意事項再填寫本頁} -訂_ 經濟部中央標準局員工消費合作社印製 A7 __— —__B7五、發明説明（6 ) fij = Σ Ρ/ο^Β/ω) 其中Pi(c〇)是分析框"i "在頻率ω時之能譜値，Β』(ω)是帶通 濾波器加權函數或;慮波器在頻率ω時之頻率響應，Σ代表 所有頻率ω之總合運算，且fij是分析框"〖，，及帶通濾波器 "j ··之帶通濾波器輸出。在各分析框及各帶通滤 波器"j"(〇Sl^n)之所有帶通濾波器輸出計算完成後，特徵 向量裝置110以取各帶通濾波器fij之對數計算取樣聲音信號 之特徵値(圖3之340)。結果可表示成圖5中所述之以 "i "分析框與"j "帶通濾波器建構具有ηχιη階之矩陣。所有 分析框内特徵値Vil到Vijn形成單一特徵向量（如Vii到外，項目 510) ’且所有分析框形成許多取樣聲音信號之特徵向 量0 一旦分析框"i"=l到"n "之大量的特徵向量被計算出，連 接或併入特徵向量裝置之最小/最大裝置12〇(圖丨）檢視帶 通滤波器"j ·'頻帶内所有特徵値且找出所有分析框❶^丨^頻 帶” j "之最小（min』）特徵値及最大（maXj)特徵値（圖2之 220)。這些最小及最大値使用於決定正規特徵値"v〜,，。 圖1之正規器130連接最小/最大裝置120和特徵向量裝置 110。正規器130正規化各特徵値越過一頻帶或帶通濾波器 ·· j "帶通濾波器之最小及最大特徵値以決定正規特徵値 "v〜"（圖2之230) 〇正規方程式爲·· (請先閱讀背面之注意事項再填寫本頁) 裝· 訂 • 1— I n n m n 本紙張尺度適用中國國家操準（CNS ) A4規格（210 X 297公釐）Pi (k) = s ^ k)-Si (kl) where Si (k) is the analysis frame " i " sampling of the sound signal at position k, is the analysis frame " i " time in front position " k ·〗 The sound signal of " is sampled, and still ⑻ is the analysis frame " i " and the emphasized sound signal is sampled before position k. One skilled in the art of voice recognition will recognize that other pre-emphasis filters may be used. This paper scale is applicable to the Chinese National Standard Falcon (CNS) 8.4 specifications (21〇297 mm j printed by the Ministry of Economy Central Standards Bureau employee consumer cooperatives A7 '_______B7 V. Invention description (5) ~ ~ ~~~ 2. Each analysis The pre-framed emphasized sound signal samples are band-pass filtered by a filter that serially converts different frequency bands. This filter can be applied to any calculation method that you want in the time or frequency domain. In the best specific example, the wave filter It is used in the frequency domain. However, first, the energy spectrum of the pre-emphasis sound signal sampling in the analysis frame must be calculated (320 in Figure 3). The energy spectrum is obtained from: a. The multiplication of the pre-emphasis sound signal sampling in the analysis frame Sampling of the upper trigger pulse function or weighting function. Any trigger pulse function may be applied. To explain the present invention, d can assume a simple rectangular trigger pulse (the trigger pulse value for all samples is 1.0). B. Calculate each trigger pulse The Fourier transform of the pre-emphasized sound signal samples in the analysis frame. C. Fourier transform squared to obtain the energy spectrum value. After the spectrum value is determined, the bandpass filter is applied to the frequency by the filter weighted value of each energy spectrum value. Domain (330 in Fig. 3). Although many filter weighting functions may be used in the band-pass filter, the best specific example incorporates the rising cosine weighting cross section seen in Fig. 4. Fig. 4 shows the use of the rising cosine cross section 41. In the above energy spectrum 400. In the best embodiment of the present invention, the frequency band of each band-pass filter or rising cosine cross-section 41o according to the Maier or Barker scale's frequency response close to the frequency response of the human ear is plotted along the frequency axis. The frequency band of the band-pass filter (rising cosine cross-section 41〇) tends to be linearly spaced from 0 to 1 kHz, and is l0grithmically separated by more than 1 kHz. The wave spacings other than those defined by the best specific examples are also It can be used. As seen in the best specific example of Figure 4, the bandpass filter or rising cosine cross section 41 is overlapped. The calculation of the output of the band pass filter or rising cosine cross section 410 is based on: This paper scale applies the Chinese national standard ( CNS) M specification (210x297 mm) (please read the notes on the back before filling in this page) -Subscribe _ Printed by the Ministry of Economy Central Standards Bureau Employee Consumer Cooperative A7 __— —__ B7 V. Description of the invention (6) fij = Σ Ρ / ^ Β / ω) where Pi (c〇) is the analysis frame " i " energy spectrum value at frequency ω, Β 』(ω) is the bandpass filter weighting function or; when the wave filter is at frequency ω Frequency response, Σ represents the total operation of all frequencies ω, and fij is the output of the band-pass filter of the analysis box " 〖, and band-pass filter " j. In each analysis box and each band-pass filter " j " (〇Sl ^ n) after all output calculations of the bandpass filter are completed, the feature vector device 110 takes the logarithm of each bandpass filter fij to calculate the feature value of the sampled sound signal (340 of FIG. 3). The result can be expressed as the matrix of ηχιη order constructed with " i " analysis frame and " j " band-pass filter as described in FIG. 5. All the feature values in the analysis frame Vil to Vijn form a single feature vector (such as Vii to the outside, item 510) 'and all analysis frames form a feature vector of many sampled sound signals 0 Once the analysis frame " i " = l to " n " A large number of feature vectors are calculated, connected or merged into the minimum / maximum device of the feature vector device 12〇 (Figure 丨) to view the band-pass filter " j · 'all the feature values in the band and find all the analysis frames ❶ ^ 丨 ^ The minimum (min) characteristic value and the maximum (maXj) characteristic value of the frequency band "j" (Figure 2 of 220). These minimum and maximum values are used to determine the normal characteristic value "v ~,". Figure 1 The regularizer 130 connects the minimum / maximum device 120 and the feature vector device 110. The regularizer 130 normalizes each feature value across a frequency band or bandpass filter ... j " band pass filter minimum and maximum feature values to determine the regularity Characteristic value " v ~ " (230 in Figure 2) ○ The normal equation is (please read the precautions on the back before filling in this page) Binding · Order • 1—I nnmn This paper size is applicable to the Chinese national standards ( CNS) A4 specification (210 X 297 PCT)
五、發明説明（7 經濟部中央標準局員工消費合作社印製 = (Vij - minj)/(maxj - minj) 其中ν〜ϋ是其中之—正規特徵値，％是其中之—特徵値， minj疋"j "頻帶之最小特徵値，且瓜％是"产，，頻帶之最大特 徵値。 正規化方法的結果可表示成圖6所描述之矩陣。圖6之各 分析框"i"代表一正規特徵向量（61〇)。 圖1 t比對器140連接正規器13〇且將正規特徵向量與樣板 特徵向量做比對以決定出冑一個樣板特徵向量最類似此正 規特徵向量。表不片語或命令的樣板特徵向量集合儲存於 樣板特徵向量庫150内。比對器14〇依次比對來自正規器13〇 之正規特徵向量及樣板特徵向量庫15〇内各樣板特徵向量 (圖2之240)且決定出哪—個樣板特徵向量集合最類似此正 規特徵向量（250)。這將以計算正規特徵向量與各樣板特 徵向量集合間之間距來完成。具有最小間距之樣板特徵向 量集合被決定爲最類似此正規特徵向量的一個。圖丨之比 對器140由最類似（具有最小距離標度）正規特徵向量（圖2 的250)的樣板特徵向量庫15〇中輸出一最吻合之樣板特徵 向量集合。 有幾種有名的方法因此許多正規特徵向量可與樣板特徵 向量比對找出最吻合的。研究使用本發明最佳具體實例顯 示在機動時間彎法中比對許多正規特徵向量與樣板特徵向 量產生最佳結果。 如早先所提的，當本發明使用依賴話筒，小語彙聲音辨 識系統是非常健全的且增加在高度或易變噪音環境自不堪 (請先閲讀背面之注意事項再填寫本頁) 策· -訂 10· 五、發明説明（8 A7 B7 使用的準確率辨識準確性到非,高的準確率 音 聲 同 不 多 月 許吏 在i 用類 使種 能同 可不 明有 發所 本。 的 發 本 期 預 均 認 確 該 應 是 統 系 識。 辨明 ：--------{裝— (請先閱讀背面之注意事項再填寫本頁) 訂 經濟部中央標準局員工消費合作社印製 本紙張尺度適用中國國家標準（CNS ) A4規格（210X297公釐）V. Description of the invention (7 Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs = (Vij-minj) / (maxj-minj) where ν ~ ϋ is one of them-regular characteristic value,% is one of them-characteristic value, minj " j " The minimum characteristic value of the frequency band, and the% is the maximum characteristic value of the frequency band. The result of the normalization method can be expressed as the matrix described in FIG. 6. The analysis boxes of FIG. 6 " i "; Represents a normal feature vector (61〇). Figure 1 t comparer 140 is connected to the regularizer 13〇 and the normal feature vector is compared with the model feature vector to determine that a model feature vector is most similar to this regular feature vector. The set of template feature vectors representing phrases or commands is stored in the template feature vector library 150. The comparator 14〇 sequentially compares the normal feature vectors from the regularizer 13〇 and the template feature vectors in the template feature vector library 15 ( (240 of Fig. 2) and determine which set of model feature vectors is most similar to the regular feature vector (250). This will be done by calculating the distance between the regular feature vector and each set of model feature vectors. The minimum spacing The set of template feature vectors is determined to be the one that is most similar to the normal feature vector. The comparator 140 in FIG. 1 is composed of a model feature vector library 15 that is the most similar (with the smallest distance scale) normal feature vector (250 in FIG. 2). Output a set of best-fitting model feature vectors. There are several well-known methods. Therefore, many regular feature vectors can be compared with the model feature vectors to find the best match. The best specific example of using the present invention is shown in the maneuver time bending method. It produces the best results for many regular feature vectors and model feature vectors. As mentioned earlier, when the present invention uses a microphone, the small vocabulary voice recognition system is very robust and increases in height or variable noise environment is unbearable (please first Read the precautions on the back and then fill out this page) Policy · -Subscribe 10 · V. Description of the invention (8 A7 B7 The accuracy rate used to identify the accuracy is not correct, the high accuracy rate sounds are not many months Xu Li is used in i class It can be used to make the species unclear. The current issue of this book presupposes that it should be the system identification. Identify: -------- {install— (please read the back side first Please pay attention to this page and then fill out this page) Order Printed by the Employee Consumer Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs This paper standard applies to China National Standard (CNS) A4 (210X297mm)