4 53 04 υ A7 B7 五、發明説明（ <發明背景> 拜電腦多媒體發達之所賜，數位語音在電腦上的 ， 廣泛’尤其是電腦網路的蓬勃發展，使得語音與網路的^^ 目前最新的發展趨勢，例如當前最流行的網路電話 ，都是最㈣例子。 ⑺日冊寺 語音傳遞的目的在於溝通，因此，其安全性是個相春重要的 考量，尤其是當傳遞的内容牽涉到個人隱私或商業機密^，更不 容!If不法人士所竊聽。因此’我們必須使用某種方法來保護它 。最簡單㈣式就是使㈣統的文字㈣加料直雜語音資料 做加密’顧這種加密法本身具備相#高的安全性，但^一般 的語音資料都相當龐大，如果使用這種加密絲加密，會浪費相 當多的時間。而且’語音資料的特性和—般的文字資料也有相當 大的出入’對於語音訊號而言’些許的失真並不會對人類的聽覺 造成太大㈣響’也由於這個緣故’我舰不需特別要求解密之 後的語音職-定得和原始的語音訊m只要接收方能 辨識出解密後之語音訊號所代表的含意即可。 根據這項特性，學者們特別提出了一些針對語音訊號所設計 出來的加密系統，這些加密系統藉由擾亂頻域或時域之下的語音 汛號而達到語音加密的目的，而除了會對原本的語音訊號造成少 许失真外，它們比傳統的加密系統擁有更快的加密速度以及更好 的雜訊抵抗能力。 請 先 閱 ★ 背 之 注 意 事 項 再 填 寫 本 頁 裝4 53 04 υ A7 B7 V. Description of the invention (< Background of invention > Thanks to the development of computer multimedia, digital voice is widely used on computers, especially the booming development of computer networks, which makes voice and the Internet ^ ^ The latest development trends, such as the most popular Internet telephones, are the best examples. The purpose of voice transmission in Nichiriji Temple is communication. Therefore, its security is an important consideration, especially when it is transmitted. The content involves personal privacy or trade secrets ^, even more intolerable! If it is eavesdropped by an unlawful person. So 'we must use some method to protect it. The simplest way is to make the normal text plus the mixed voice data for encryption' This encryption method itself has relatively high security, but the general voice data is quite large. If you use this kind of encryption wire to encrypt, it will waste a lot of time. And the characteristics of voice data and ordinary text There is also a considerable discrepancy in the data. 'For voice signals,' a little distortion does not cause too much noise in human hearing. 'For this reason, our ship There is no special requirement for the voice post-decryption and original voice signal m after decryption, as long as the receiver can identify the meaning represented by the decrypted voice signal. Based on this feature, scholars have proposed some specific voice signals Designed encryption systems. These encryption systems achieve the purpose of voice encryption by disturbing the voice flood numbers in the frequency or time domain. In addition to causing a little distortion to the original voice signal, they are better than traditional encryption systems. Have faster encryption speed and better noise resistance. Please read the note before filling in this page to install
tT 經濟部中央標4局員工消贽合作钍印製 本'.氏張尺度制巾賴家轉（CN.S ) A4祕（2I0X297公竣） 45304 6 * 經满部中央標準扃員工消费合作社印褽 A7 B7 五、發明説明（2 ) 然而，經由這些語音加密系統所產生出來的加密結果仍舊有 其缺點存在，因為這些加密後的訊號所呈現出來的都是一些雜亂 和無意義的訊號，這樣的結果雖然可以保護我們的機密語音，但 同樣也會惹來不法者的注意。而當這些不法者截獲此機密語音並 且發現它已經被加密過，直覺的會認為這是個相當重要的語音訊 息。於是，他們便會想盡辦法去破解它，即使他們無法破解，也 可能會去破壞或變造它，使接收者無法順利的接收此機密語音。 因此，我們提出一個新的語音加密系統來避免這種不法者的 直接攻擊。在這個加密系統中，機密語音是藏在一段普通語音中 ’此普通語音是一段無足輕重而且可以公開的語音，即使它被不 法人士所操取’不法人士很可能相信這段語音就是傳送者所要傳 送的原始語音’因此沒有理由去破解它〇同時，由於此語音並不 具特別的意義，因此他們也沒理由會去破壞或偽造它，故接收者 可以順利的收到這段語音並還原它《此外，由於在隱藏之前，我 們還將此機密語音以對稱式加密系統加密過，這樣，即使不法者 取得此機密語音，由於沒有解密的金匙，故亦無法將它還原。因 此我們所提出的語音加密系統較之前所提出的方法更加安全。 1¾衣------ΐτ------ (請先閱讀背面之注意事項再填寫本頁} 4^3〇4 6 A7 五、發明說明（3 ) <發明目的> 本發明之目的在於提供一植基於次頻帶分割之語音偽裝技術 ，’、、主要，利用頻帶分割法將一般語音的頻帶分割成數個次頻帶 ^將機密語音加紐隱藏於較高_次鮮巾以蒙蔽不法者， 唯有合法者之麵金匙可將此機密語音取出。 本發明之另一目的在於提供一植基於次頻帶分 ，，藉由將機密語音壓縮並隱藏於一般語音“壞力的 局頻次頻帶中’可有效降低整體語音的資料量而 音的壓縮效率_亦不至於造祕重的失真。㈣大从^ --------—t------IT------線 {請先聞讀背面之注Jt-e事項再填寫本頁) 經消部中央榡準局負工消費合作社印裝 本紙張尺度適用悄财麵（cns ) Λ4^ (21Qx297公楚） Α7 Β7 五、發明説明（4 經濟部中央標準局負工消费合作社印裝 <發明之詳細說明> 我們所提出來的方法主要是架構在次頻帶切割（^^^4 Division)的技術上，該技術藉由將語音頻帶切割成數個頻寬較小 的次頻帶’可有效降低各次頻帶取樣間的相關性，並增進編碼的 效率。將此技術首先應用在語音編碼上的是由Cr〇c}liere、 和Flanagan等三人於1976年所提出來的分頻編碼法。近年來這種 利用頻帶切割觀念的編碼技術也已漸漸成為語音編碼中的主流， 眾所熟知的MPEG音訊編碼即是' —例。 以雙次頻帶之分頻編碼法為例，其基本作法是利用高通淚波 器及低通濾波盗將輸入的語音頻帶分割成高頻和低頻兩個次頻帶 ，母個次頻帶再分開來做編碼的動作。這種分割的方法可以將原 本輸入訊號的相關性去除，而所得到的輸出訊號其資料相關性較 原訊號為低’這樣做有助於提高編碼的效率。此外，次頻帶分巧 最主要的優點是每個次頻帶的編碼可以依其特徵的不同而分別指 定不同的位元率，低頻的次頻帶由於需準確的保留語音的音調及 波形結構，因此我們以較多的位元來做編碼；而高頻的次頻帶大 部分是摩擦音或類似雜訊的聲音，因此我們只需使用少量的位元 來做編瑪。 由於分割後之低頻訊號及高頻訊號的頻寬都比原語音訊號之 頻寬為小，因此在這裡我們需先將這兩個次頻帶以原取樣頻率的 一半做次取樣，然後才進行編碼的動作。每個次頻帶的編碼是根 據人類的知覺標準來指定其編碼方法，低頻的次頻帶由於其取樣 間的累贅還很尚’故我們使用誤差訊號編碼（Differential Pulse —壯衣-- (請先閱讀背面之注意事項再填寫本頁) -T ·-* t. 私紙張尺度適用中國國家標率{ CNS > A4規格（2丨 好湞部中央標準局員工消f合作社印裝 45304 A? ___B7 五、發明説明（5 ) ~~tT The Central Bureau of the Ministry of Economic Affairs, the 4th Bureau of the People's Republic of China eliminates the cooperation and prints the printed version of the paper.'s Zhang scale towel Lai Jiazhuan (CN.S) A4 secret (2I0X297 public completion) 45304 6褽 A7 B7 V. Description of the invention (2) However, the encryption results produced by these voice encryption systems still have their shortcomings, because these encrypted signals show some cluttered and meaningless signals. Although the result can protect our confidential voice, it will also attract the wrongdoers' attention. When these criminals intercept this confidential voice and find that it has been encrypted, it is intuitively considered that this is a very important voice message. Therefore, they will try their best to crack it. Even if they cannot crack it, they may destroy or alter it, so that the receiver cannot successfully receive this confidential voice. Therefore, we propose a new speech encryption system to avoid such direct attacks from the wrongdoers. In this encryption system, the confidential voice is hidden in a common voice. 'This common voice is an insignificant and public voice, even if it is manipulated by a wrongdoer.' An unlawful person is likely to believe that this voice is what the sender wants to transmit. Therefore, there is no reason to crack the original voice. At the same time, because this voice has no special meaning, they have no reason to destroy or forge it, so the receiver can smoothly receive this voice and restore it. Because the confidential voice was encrypted with a symmetric encryption system before being hidden, so that even if an unscrupulous person obtains this confidential voice, it cannot be restored because there is no decryption key. Therefore, our proposed speech encryption system is more secure than the previously proposed method. 1¾ 衣 ------ ΐτ ------ (Please read the notes on the back before filling in this page} 4 ^ 3〇4 6 A7 V. Description of the invention (3) < Object of the invention > The present invention The purpose is to provide a voice camouflage technology based on sub-band segmentation. ',, and mainly use the frequency band division method to divide the general voice frequency band into several sub-bands. For the wrongdoer, only the legal person's face key can take out this confidential voice. Another object of the present invention is to provide a sub-band based, which compresses and hides the confidential voice in the general voice "bad power bureau". 'In the frequency band' can effectively reduce the amount of data in the overall speech and the compression efficiency of the sound _ will not create heavy distortion. ㈣ 大 从 ^ ---------- t ------ IT-- ---- Line {Please read the note on the back of Jt-e before filling out this page.) The paper size of the printed paper is applicable to the paper size (cns) Λ4 ^ (21Qx297) Chu) Α7 Β7 V. Description of the invention (4 Printed by the Consumers' Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs < Detailed description of the invention > I The method we proposed is mainly based on the technology of sub-band division (^^^ 4 Division). This technology can effectively reduce the sampling interval of each sub-band by cutting the voice band into several sub-bands with smaller bandwidth. Correlation and improve the efficiency of coding. The first application of this technology to speech coding is the frequency division coding method proposed by Croclielie, Flanagan, and others in 1976. In recent years, this use The coding technology of the band-cut concept has gradually become the mainstream in speech coding, and the well-known MPEG audio coding is the example. Taking the frequency division coding method of the double frequency band as an example, the basic method is to use a high-pass tear wave device. And low-pass filtering, the input voice frequency band is divided into two high-frequency and low-frequency sub-bands, and the mother sub-bands are then separated to perform the encoding operation. This method of division can remove the correlation of the original input signal, and therefore The obtained output signal has lower data correlation than the original signal. 'This will help improve the coding efficiency. In addition, the main advantage of sub-band division is that the coding of each sub-band can be based on its characteristics. Different bit rates are specified for different characteristics. The low-frequency sub-bands need to accurately retain the pitch and waveform structure of the speech, so we use more bits to encode; the high-frequency sub-bands are mostly fricative. Or similar noise, so we only need to use a small number of bits to edit. Since the bandwidth of the divided low-frequency and high-frequency signals is smaller than the bandwidth of the original voice signal, we need here The two sub-bands are sub-sampled at half of the original sampling frequency, and then the encoding operation is performed. The coding of each sub-band is based on the human perception standard to specify its encoding method. The burden is still too 'so we use error signal coding (Differential Pulse-strong clothing-(Please read the precautions on the back before filling out this page) -T ·-* t. Private paper scales apply Chinese national standard {CNS > A4 specification (2 丨 Goods printed by the Central Bureau of Standards of the Ministry of Consumers and Cooperatives 45304 A? ___B7 V. Description of the invention (5) ~~
Code Modulation，以下簡稱DPCM)法來做編碼，至於高頻的次頻 帶，由於取樣間的累贅已經很低，因此我們可以直接做量化編碼 (PCM編碼）。 在重建語音訊號時’每個次頻帶分別以其相對應的解碼器進 行解碼，解碼後的結果先以1 :2提高取樣率，然後通過適當的滅 波器，最後再加起來即可得到重建的語音訊號。 以上所描述的即為分頻編碼法的簡單概念與流程，我們的方 法就疋利用此編碼法來達到語音壓縮與偽裝的目的。假設現在傳 送方想要傳送一段機密語音給接收方’那麼他就必須準備另一段 不具重要性的一般語音來隱藏這段機密語音，其流程如圖一所示 ’我們以S來表示這段機密語音，以C來代表一般語音，而處理 後的結果我們稱之為偽裝語音。在將S藏入C之前，傳送方先以 下列步驟來對$做處理： 1·次頻帶分解（Subband Division) 將S以高通濾波器及低通濾波器分解成高頻和低頻兩個次 頻帶’由於語音訊號的音調及波形結構主要是由低頻次頻 帶的部分所保持，因此，為了減少藏入一般語音中的資訊 量，我們只保留低頻的次頻帶而把高頻的次頻帶捨棄掉。 接著’再將保留下來的低頻次頻帶以原訊號頻率的一半做 次取樣，我們以&來代表此次取樣後的結果。在這裡我們 可以繼續以上面的步驟重複對&進行分解及次取樣的動作 ’只要最後保留下來的低頻次頻帶&，其失真程度在我們 本紙法λ“用家標华（CNS) M規格（2!Qx297公楚）- -~~ 1¾衣------訂------^ (諳先閱讀背面之注意事項再填寫本頁} 部 中 央 標 準 扁 貝 工 消 η ίΐ 印 製 本紙張尺度適用中囤闽家標率（(：1^)^4規格（2丨0>< 297公釐） 4 53〇z?. Α7 — Β7____ 五、發明説明（6 ) 可接受的範圍之内，亦即此語音訊號所要表達的意義仍舊 可以很輕易的被我們所辨識。 2. 編碼 因為形成次頻帶的過程本身並不具有資料壓縮的效果，為 了進一步減少資訊的隱藏量，我們必須再選擇一個適當的 編碼法來對&做編碼。考慮低頻的次頻帶其取樣間的累贅 還很高，我們選擇以DPCM來做編碼，其輸出結果以$表 示之。 3. 加密 為了增加機密語音的安全性’我們有必要對$做進一步的 加密動作。加密的方法我們可以從各種對稱式的加密系統 如 Data Encryption Standard (DES)、International DataCode Modulation (hereinafter referred to as DPCM) method for coding. As for the high-frequency sub-frequency band, since the burden between samples is already low, we can directly perform quantization coding (PCM coding). When reconstructing a voice signal, 'each sub-band is decoded by its corresponding decoder, and the decoded result is first increased by 1: 2, then passed through an appropriate wave suppressor, and finally added together to obtain reconstruction. Voice signal. What is described above is the simple concept and flow of the frequency division coding method. Our method does not use this coding method to achieve the purpose of speech compression and disguise. Suppose now that the sender wants to send a confidential voice to the receiver, then he must prepare another general voice that is not important to hide this confidential voice. The flow is shown in Figure 1. 'We use S to represent this confidential voice. Speech, the general speech is represented by C, and the processed result is called camouflage speech. Before hiding S in C, the sender processes $ in the following steps: 1. Subband Division: S is decomposed into high-frequency and low-frequency sub-bands by a high-pass filter and a low-pass filter. 'Because the tone and waveform structure of the voice signal are mainly maintained by the low frequency subband, in order to reduce the amount of information hidden in general speech, we only retain the low frequency subband and discard the high frequency subband. Then, the remaining low frequency sub-band is sampled at half the original signal frequency. We use & to represent the result of this sampling. Here we can continue with the above steps to repeat the decomposition and sub-sampling action 'as long as the low-frequency sub-frequency band & is retained at the end, the degree of distortion is in our paper method λ "using standard Chinese (CNS) M specifications （2! Qx297 公 楚）--~~ 1¾ Clothing ------ Order ------ ^ (谙 Please read the precautions on the back before filling in this page} Ministry of Standard Standard Flat Shellfish Consumption The paper size of the paper is applicable to the standard of the domestic standard ((: 1 ^) ^ 4 specifications (2 丨 0 > < 297 mm) 4 53〇z ?. Α7 — Β7 ____ 5. Description of the invention (6) Acceptable Within the range, that is, the meaning of this voice signal can still be easily recognized by us. 2. Because the process of forming the sub-band does not have the effect of data compression itself, in order to further reduce the hidden amount of information, we We must choose an appropriate encoding method to encode &. Considering the low frequency sub-band, the burden between samples is still very high, we choose to use DPCM to encode, and the output result is expressed in $. 3. Encryption in order to increase Security of confidential voice 'We must For $ further encryption operation. We encryption method from a variety of symmetric encryption system such as Data Encryption Standard (DES), International Data
EncryptionAlgorithm(IDEA)等任選其中一種，並以其秘密 金匙A：來對名進行加密。加密後的結果即為我們最後要 藏入C中的機密語音資料。 關於以上對於機密語音的處理步驟，我們以圖三之流程圖 表示之，其中心為此機密語音的取樣頻率β接著，對於一 般語音C，我們以下列的步驟來處理： 1.次頻帶分解 將C以禹通滅波及低通渡波分解成高頻和低頻兩個次 頻帶’然後再以原訊號頻率的一半對這兩個次頻帶做次取 ---------—裝------訂------0 (請先閱讀背面之注意事項再填寫本頁) 4 5304 e A7 ____ B7 五、發明説明（7 ) — 樣’取樣後的結果’高頻的部分我們以Q來表示，而低頻 的部分則以CV來表示。 (請先閱讀背面之注意事項再填寫本頁) 2. 編碼及語音隱藏 由於低頻的次頻帶保留了大部分語音的特質，因此在考慮 語音隱藏時，我們以高頻的次頻帶作為資料隱藏的對象， 而對於低頻的次頻帶，我們仍舊以DPCM來做編碼，其輸 出結果以c丨表不之。至於南頻的次頻帶，由於其訊號幾乎 呈現白色雜訊（WhiteNoise)的形式而且所佔的能量百分比 非常小，因此將我們的機密語音隱藏在這個部分比較不容 易被攻擊者所察覺，同時也不會對整個語音訊號的品質產 生太大的影響。以下我們就介紹將機密語音（即治隱藏到 此高頻的次頻帶（即CA)的方法。 假設贫的長度為〜個位元組（即〜X 8個位元），而共有 〜個取樣《首先，對CA做量化（Quantization)的動作，量 化的結果以C；«表示之’由於尚未做編碼的動作，因此其取 樣數仍舊為〜個。在這裡我們定義一個新的參數 r nc~m， 其中m為用來對〜之數值編碼的位元數，此參數的意義是 當我們將萬藏入C〖後’最多會更動到q的取樣之位元數 目。接著我們先將〜的數值嵌入C;前w個取樣的最後一 個位元中，然後再將另依序均勻的嵌入q其餘取樣的最後 _，其輪出結果以π袅示之，诘_ r'f'會和有 本紙张尺度適用中國囤家樣準（〇灿）/\4規格（210+/297公釐} 9 A7EncryptionAlgorithm (IDEA), etc., and use its secret key A: to encrypt the name. The encrypted result is the confidential voice data that we will finally hide in C. Regarding the above processing steps for confidential speech, we use the flow chart in Figure 3 to represent the sampling frequency β of the confidential speech. Next, for the general speech C, we process the following steps: 1. Sub-band decomposition will C uses Yutong annihilation wave and low-pass crossing wave to decompose into two high-frequency and low-frequency sub-bands, and then uses the half of the original signal frequency to take the two sub-bands. ---- Order ------ 0 (Please read the notes on the back before filling out this page) 4 5304 e A7 ____ B7 V. Description of the invention (7) — Sample 'Results after sampling' high frequency part We use Q to represent it, while the low frequency part is represented by CV. (Please read the notes on the back before filling out this page) 2. Coding and voice concealment Because the low-frequency sub-band retains most of the characteristics of speech, we consider the high-frequency sub-band as the data concealment when considering voice concealment. Object, and for the low frequency sub-band, we still use DPCM to encode, and the output result is not as good as c 丨. As for the sub-band of the south frequency, because its signal is almost in the form of White Noise and its energy percentage is very small, hiding our confidential voice in this part is relatively difficult for attackers to detect, and also Does not have much impact on the quality of the entire voice signal. In the following, we will introduce the method of concealing confidential speech (ie, hiding the sub-band (CA) to this high frequency). Assume that the length of the lean is ~ bytes (ie ~ X 8 bits), and there are ~ samples "First, perform a quantization action on CA. The result of quantization is C;« indicates that 'because the encoding action has not yet been performed, so the number of samples is still ~. Here we define a new parameter r nc ~ m, where m is the number of bits used to encode the value of ~, the meaning of this parameter is that when we hide 10,000 in C, the number of samples will be changed to q at most. Then we first The value is embedded in C; the last bit of the first w samples, and then the other _ of the remaining samples of q are evenly and sequentially embedded. The result of the round-out is shown by π 袅, 诘 _ r'f ' This paper size applies to Chinese storehouse samples (〇 Chan) / \ 4 specifications (210 + / 297 mm) 9 A7
4 53 〇 4 p 五、發明説明（8 ) 少其差異有限’故當我們還原回-般語音時 '、’會產生太大的誤差。待所有資料隱藏完畢之後，便 :壓:Ϊ進::了:動在14裡我們採用無失真的的資 料壓縣’如Huffman編碼法，來對q做編碼；如此4 53 〇 4 p V. Description of the invention (8) The difference is limited, so when we restore the normal speech, ',' will produce too much error. After all the data is hidden, then: Press: 压 进 :: 了: In the 14th, we use the distortion-free data pressure county ’such as Huffman coding method to encode q; so
們解碼時才能夠準確的還原回〇;,。編石馬之後的結果則^ 表示之。 A 最後’傳送方需將巧金匙&透過安全的管道傳送哈接收 方，然後再將G;和3傳送給減脚可。圖四為二上對 於-般語音的處理流程示意圖，其中Fc為一般語音的取樣 頻率。 重建機密語音的方法如圖二所示，接收方必需擁有合法的 秘密金匙，並以之對所接收到的偽裝語音進行解密，最後 還原出機密語音’其詳細的過程描述如下。 我們先將C以其相對應之解碼器解瑪得到C ,接著從 的前m個取樣取出最後一個位元，以還原回〜之數值，再 由〜算出r，然後我們從g取出剩餘取樣的最後固位元 以還原出贫。接下來，以合法的秘密金匙A對另進行解密 得到S) ’再以DPCM解碼器對s丨做解碼；由於DPCM是 一種會失真的壓縮法’故此時所得到的解碼結果與原先的 &有些許差異’但此差異極為有限，以人類的聽覺很難分 辨，在此我們用毛來表示$解碼後的結果。最後只要將-以1 :2提高取樣率並通過適當的遽波器即可得到還原後 之機密語音S，此機密語音與原始的機密語音非常類似。 4、..·(^尺度適用中国國家標準（CNS ) Λ4規格（2|〇χ297公嫠） 10 --------1#衣------訂------Μ (請先閱讀背面之注意事項再填寫本頁〕 經濟部中央標準局男工消費合作社印繁 ^304 6 A7 B7 經漪部中夾標孳局員工消费合作社印製 五、發明説明（9 為了證明我們所提出來的方法確實可行，我們進行以下兩個 實驗，第一個實驗疋選擇一段機密語音並利用我們所提出來的方 法將它藏人另-段-般語音中’在這個實驗中我們所使用的機密 語音長度為1.48秒，其實際的龍大小為65，174位元組，而一般 語音的長度為1.96秒，其資料大小為86,216位元組；它們的取樣 頻率皆為22,050取樣/秒，而每個取樣皆使用丨6個位元來編碼: 為了取得對照的數據，在實驗之前我們先以標準的適應性DpcM (以下簡稱ADPCM)語音壓縮法，分別對一般語音及機密語音做 壓縮，其壓縮比皆為固定的4 : 1。接著機密語音以我們所提出來 的方法做一次的次頻帶分解，然後再編碼、加密後，所得到的語 音資料大小為9,719位元組，故而其壓縮比為6 71 :丨。至於一般 語音，我們單獨地以分頻編碼法做語音壓縮後所得到的輸出大小 為21，458位元組，壓縮比為4.02 : 1 ;如果進一步以我們的方法 做次頻帶分解、編碼，然後藏入機密語音後，所得到的輪出大小 為23,448位元組’壓縮比為3·68 : 1。由於機密語音已藏入一般 語音令，故整體的壓縮比為（65，174+86,216) : 23,448 = 6.46 : 1。 由以上的實驗結果可知，雖然一般語音經藏入機密語音後，其個 別的壓縮比降低了少許，但考慮機密語音消失之後的整體壓縮比 ’其效率仍舊比標準的語音壓縮高出甚多。 為了便於瞭解失真的情形，我們引用SNR (Signal to Noise Ratio)來判斷處理前後的差異程度。其計算方式為： 顧=10 X log 本紙張尺度適财關家料（CNS) Λ4規格（21()>< 297公麥） --------—裝------訂------妹 (諸先閱讀背面之注意事項再填寫本頁) 好"部中火標隼局員工消贫合作社印^ 453〇46 — A7 - __ B 7 五、發明説明（10 ) ‘ 其中，κ⑺為輪入訊號，印）為輪出訊號，rt則為取樣數目，而 SNR值愈高，表示處理前後的差異程度愈小，亦即經處理過後的 訊號失真愈少。 在本實驗中，一般語音和機密語音以ADPCM編碼法壓縮後 ’其SNR值分別為30.332 dB及30.539，而一般語音以本方法做 語音壓縮後，其SNR值為33.754 dB，在進_步藏入機密語音後 ，其SNR值為32.512 dB，這兩個結果都比標準的語音壓縮還要 好。圖七所示為本實驗所使用的原始一般語音訊號的波形，圖八 所示為其經壓縮並藏入機密語音後之一般語音訊號波形，圖九則 為將兩訊號相減所形成的誤差波形，用來表示處理前後的訊號差 異程度。由以上結果可知，一般語音在藏入機密語音後所引起的 失真程度並不大，一般人很難察覺這段語音已多藏匿了另外一段 語音。另一方面，隱藏於一般語音中的機密語音經還原之後，其 SNR值為31.149dB，和標準的語音壓縮比起來，這種品質亦遠高 於我們可以接受的範圍。圖十所示為本實驗所使用的原始機密語 音訊號的波形，圖Η * —所示為其經壓縮隱藏並還原之後的語音訊 號波形，圖十二則為兩訊號相減所形成的誤差波形。 在第二個實驗裡’我們令機密語音維持不變，因此其壓縮比 和失真程度與上一個實驗相同。另一方面，一般語音則加長為4.1 秒’資料大小變為180,628位元組，將它以ADPCM編碼後，其 壓縮比為4 : 1，SNR值為30.332 dB ;另外，將其以我們所提之 方法做語音壓縮及藏入機密語音後，其大小變為43,〇7〇位元組， 壓縮比為4.19 ·’ 1 ’整體的壓縮比為（65，174+180,628) : 43,070 = 本紙乐尺度適用中國SJ家標準_(_CNS〉A4規格（21〇x 297公釐）"~~'~~' ~~' 12 I- m^i I -- .^1^^g -- I— -- ^1. --iJI - F I I I I (請先閱讀背面之注意事項再填寫本頁) 453〇4r A7 ----____B7 五、發明説明（11 ) 5,70 : 1，至於其SNR值則為33.071 dB，此值亦遠高於標準的語 曰壓縮法。圖十三所示為本實驗所使用的原始一般語音訊號的波 形’圖十四所示為其經壓縮並藏入機密語音後之一般語音訊號波 形’圖十五則為將兩訊號相減所形成的誤差波形。由此實驗結果 了知，^我們增加一般δ吾音的大小後，雖然會使壓縮效率下降， 但卻可提高偽裝後的語音品質。 ----— I I I ! n I n ----丁.----- n I 表 U5 ，-0 分 (諸先聞讀背面之注意事項再填寫本頁) 經濟部中央榡车局員工消贽合作钍印製 Μ 一度 尺 ^ 紙They can be accurately restored back to 0 ;, when they are decoded. The result after weaving a stone horse is ^. A Finally, the sender needs to send the smart gold key & the receiver through a secure channel, and then G and 3 to the minus foot. Figure 4 is a schematic diagram of the processing of normal speech on the second, where Fc is the sampling frequency of general speech. The method for reconstructing the confidential voice is shown in Figure 2. The receiver must have a legitimate secret key and use it to decrypt the received masqueraded voice. Finally, the confidential voice is restored. The detailed process is described below. We first decompose C with its corresponding decoder to get C, then take the last bit from the first m samples to restore the value of ~, and then calculate r from ~, then we take the remaining samples from g Finally retain the element to restore poverty. Next, decrypt the other with the legitimate secret key A to get S) 'and then decode s 丨 with a DPCM decoder; because DPCM is a distortion compression method', the decoding result obtained at this time is the same as the original & amp There are a few differences', but this difference is extremely limited, and it is difficult to distinguish with human hearing. Here we use Mao to represent the result after $ decoding. In the end, as long as the sampling rate is increased by 1: 2 and passed through an appropriate wave filter, the restored confidential voice S can be obtained. This confidential voice is very similar to the original confidential voice. 4. .. (^ scales apply Chinese National Standard (CNS) Λ4 specifications (2 | 〇χ297 公 嫠) 10 -------- 1 # 衣 -------- Order ------ Μ (Please read the notes on the back before filling out this page) Printed by Male Consumers Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs ^ 304 6 A7 B7 Printed by the Ministry of Economic Affairs of the Bureau of Consumers ’Cooperatives of the Bureau. To prove that the method we have proposed is indeed feasible, we perform the following two experiments. The first experiment is to select a confidential speech and use it to hide it in another segment-like speech. In this experiment, The length of the secret speech we use is 1.48 seconds, and the actual dragon size is 65,174 bytes. The length of the general speech is 1.96 seconds, and the data size is 86,216 bytes. Their sampling frequency is 22,050 samples. / S, and each sample is encoded using 6 bits: In order to obtain the control data, we first use the standard adaptive DpcM (hereinafter referred to as ADPCM) speech compression method before the experiment, respectively for general speech and confidential speech. Compression, the compression ratio is fixed 4: 1. Then The dense speech is decomposed once by the method we proposed, and then encoded and encrypted. The size of the obtained speech data is 9,719 bytes, so its compression ratio is 6 71: 丨. As for general speech, we The output size after speech compression using frequency division coding method alone is 21,458 bytes, and the compression ratio is 4.02: 1; if we further perform sub-band decomposition and encoding by our method, and then hide the confidential speech The resulting round-out size is 23,448 bytes. The compression ratio is 3.68: 1. Since the confidential voice has been hidden in the general voice order, the overall compression ratio is (65,174 + 86,216): 23,448 = 6.46: 1. From the above experimental results, we can see that although the general compression rate of individual voices is slightly reduced after being hidden in confidential voices, the overall compression rate after the disappearance of confidential voices is still more efficient than standard voice compression. In order to understand the situation of distortion, we quote the SNR (Signal to Noise Ratio) to judge the difference before and after processing. The calculation method is: Gu = 10 X log Choi Guan Family Materials (CNS) Λ4 Specification (21 () > < 297g) ---------- install -------- order ------ girl Please fill in this page before filling in the details) Good " Institute of Anti-Poverty Cooperatives of the Ministry of Fire Standards Bureau ^ 453〇46 — A7-__ B 7 V. Description of the invention (10) 'Among them, κ⑺ is a rotation signal, printed) For the round signal, rt is the number of samples, and the higher the SNR value, the smaller the difference between before and after processing, that is, the less distortion of the signal after processing. In this experiment, the general speech and confidential speech are compressed by ADPCM coding, and their SNR values are 30.332 dB and 30.539, respectively. After general speech is compressed by this method, the SNR value is 33.754 dB. After entering the confidential voice, the SNR value is 32.512 dB. Both of these results are better than the standard voice compression. Figure 7 shows the waveform of the original general voice signal used in the experiment. Figure 8 shows the general voice signal waveform after it has been compressed and hidden in the confidential voice. Figure 9 shows the error formed by subtracting the two signals. The waveform is used to indicate the degree of signal difference before and after processing. It can be seen from the above results that the distortion caused by general speech after hidden secret speech is not great, and it is difficult for ordinary people to perceive that this speech has hidden another speech. On the other hand, after the confidential speech hidden in general speech is restored, its SNR value is 31.149dB. Compared with standard speech compression, this quality is far higher than our acceptable range. Figure 10 shows the waveform of the original confidential voice signal used in the experiment. Figure Η *-shows the compressed and hidden voice signal waveform, and Figure 12 shows the error waveform formed by the subtraction of the two signals. . In the second experiment, we kept the confidential speech unchanged, so its compression ratio and distortion were the same as in the previous experiment. On the other hand, the general voice is lengthened to 4.1 seconds. The data size becomes 180,628 bytes. After encoding it with ADPCM, the compression ratio is 4: 1, and the SNR is 30.332 dB. After the method is used to compress the voice and hide the confidential voice, its size becomes 43, 〇70 bytes, and the compression ratio is 4.19. The overall compression ratio is (65,174 + 180,628): 43,070 = Ben Zhile Standards apply to Chinese SJ standards _ (_ CNS> A4 specification (21〇x 297 mm) " ~~ '~~' ~~ '12 I- m ^ i I-. ^ 1 ^^ g-I— -^ 1. --IJI-FIIII (Please read the precautions on the back before filling this page) 453〇4r A7 ----____ B7 V. Description of the invention (11) 5,70: 1, as for the SNR value The value is 33.071 dB, which is much higher than the standard language compression method. Figure 13 shows the waveform of the original general voice signal used in the experiment. Figure 14 shows the compressed and hidden secret voice. The general voice signal waveform 'Figure 15 is the error waveform formed by subtracting the two signals. From the experimental results, we know that after we increase the size of the general delta voice, It reduces the compression efficiency, but it can improve the voice quality after disguise. ----— III! N I n ---- 丁 .----- n I Table U5, -0 points Please note this page before filling out this page.) The Ministry of Economic Affairs, Central Vehicle Administration Bureau, Consumers' cooperation, printed M degree ruler ^ Paper
Ns C 準 一標 跡 公 五 、發明説明（12 ) A7 B7 <安全性之分析> 在我們所提^來的方法巾，機密語音是隱藏在-般語音中才 進行傳送的，而在實際的應用上，此—般語音是—段不具重要性 而且可以公開的語音。當我們將機密語音藏人—般語音後，只要 ”失真不太大’則&仍然是—段具有意義的語音。即使不法人士 ，得此語音，他們也不祕這是—種齡後的語音，故不會去嘗 "式破解匕’也沒有動機去破壞它^故機密語音不會有被竊聽的危 險，而且可以順利的傳送給收方。 另一方面，假設不法者已經知道此一般語音只是一種偽裝， 於是他們可以從一般語音中取出加密過的機密語音，並且設法對 機密語音進行解密_作。在__下，我們考慮三種可能的 攻擊法： 毯文攻擊法（Ciphertext-Only Attack):此攻擊假設不法者只擁 有加密過的機密語音，然後試圖從這個語音解密出原始的機 密語音。由於此機密語音已被對稱式加密系統加密過，在沒 有解密金匙的情況下，不法者無法直接獲取此機密語音的原 始内容。假設用來加密機密語音的對稱式密碼系統是安全的 ，那麼，除了使用暴力法（Brute-force)外’我們無法找到更 好的攻擊方法來取得解密用的金匙。以金匙長度為128位元 的密碼系統為例，它的金匙組合共有2nS種，假設不法者使 用一部每秒可驗證4〇〇,〇〇〇,〇〇〇把金匙的電腦來猜測解密金匙 ’他們仍然得花 2丨28/(4χ 1〇8Χ6〇Χ6〇Χ24Χ365)#2 7χ1〇22年才能 ^紙張纽適國國家標準（CNS ) ( 210X297公釐) --- 請 先 閱 背 面 意 事 項 再 填 寫 本 頁 裝 訂 經濟部中央標準局貝工消費合作社印裝 14 3 5 4 〇4 A7 —一_ _B7______ 五、發明説明（13 ) 檢查完畢’這樣的時間花費實在太鉅大了，因此我們的語音 加密系統可以完全抵禦密文攻擊法的攻擊。 2. 已知明文攻擊法（Known-Plaintext Attack):此法假設攻擊者擁 有一些解密前及解密後的機密語音組合。在這種情形下，攻 擊者有可能經由分析解密前及解密後的機密語音而求出解密 金匙。假如傳送者下次仍舊使用同一把金匙來對機密語音加 密’那麼攻擊者便可利用此金匙正確的解出其機密語音。為 了防止這種攻擊法，我們每次都使用不同的金匙來加密機密 語音，這樣即使其中一次的秘密金匙被不法者所擷取，他們 仍舊無法利用此秘密金匙來對下一個機密語音進行解密，故 在此情形下，我們的加密系統可以抵抗明文攻擊法的攻擊。 ----------种衣---„---Γίΐτ------^ (請先閱讀背面之注意事項/填寫本頁) 經濟部中央標準局貝工消費合作社印袈 I·1 表紙張尺度適用中國國家橾準（CNS ) A1 2 3 4规格（21〇Χ295公釐） 15 1 . 選擇文攻擊（Chosen-Text Attack):此法假設攻擊者不但擁有 2 一些解密前及解密後的機密語音組合，同時對於此機密語音 3 還有選擇或控制的能力。和明文攻擊法一樣，攻擊者亦有可 4 能經由分析解密前及解密後的機密語音而求出解密金匙；但 5 只要每次加密的金匙都不同，我們的系統便可抵抗這種攻擊 453〇46 453〇46 紂濟部中央標準局員工消費合作社印奴 A7 一 ~ ~ ' B7 五、發明説明（14 ) "一 <圖式說明> 圖―：本發明語音驗之_方加密、示意圖。 圖二:本發明語音餘之触方解密結構示意圖。 圖三：本發明於傳送方之機密語音處減程示意圖。 圖四.本發明於傳送方之—般語音處理流程示意圖。 圖五’ f明於接收方之-般語音解雜及隱難料#|取流程示意 圖六··本㈣於接收方之_語音娜流程示意圖。 圖七’本發明於實驗-中所使用的原始—般語音之波形示意圖。 圖八·本發明於實驗—所求得之偽裝語音的波形示意圖。 圖九：本發明於實驗一中一般語音與偽裝語音的誤差波形示意圖。 圖十.本發明於實驗一所使用的原始機密語音之波形示意圖。 圖十一：念發明於實驗一中從偽裝語音所取出之機密語音波形示意 圖0 圖十一：本發明於實驗一中機密語音經處理前後的誤差波形示意圖 0 圖十三：本發明於實驗二中所使用的原始一般語音之波形示意圖。 圖十四：本發明於實驗二所求得之偽裝語音的波形示意圖。 圖十五：本發明於實驗二中一般語音與偽裝語音的誤差波形示意圖 本紙張尺度適财咖家標準（cns )从規格（训^297公着） (請先閱讀背面之注意事項苒填寫本頁) ,裝- -i— m 16Ns C quasi-one mark five. Description of the invention (12) A7 B7 < Analysis of security > In the method we mentioned, confidential voice is transmitted in hidden voice, and in In practical applications, this—the general voice is—a segment of voice that is not important and can be made public. When we hide confidential voices—normal voices, as long as the “distortion is not too big” then & is still a meaningful voice. Even unscrupulous people get this voice, they are not secretive. This is the post-age Voice, so I wo n’t try "style cracking dagger" and have no motivation to destroy it ^ Therefore, confidential voice will not be in danger of eavesdropping, and can be smoothly transmitted to the receiver. On the other hand, it is assumed that the wrongdoer already knows this General voice is just a disguise, so they can take the encrypted secret voice from the general voice and try to decrypt the secret voice. Under __, we consider three possible attack methods: Ciphertext- Only Attack): This attack assumes that the criminal has only the encrypted secret voice, and then attempts to decrypt the original secret voice from this voice. Since this secret voice has been encrypted by the symmetric encryption system, there is no decryption key , The criminal cannot directly obtain the original content of this secret voice. Assuming that the symmetric cryptosystem used to encrypt the secret voice is secure, then In addition to using Brute-force, 'We can't find a better attack method to obtain the decryption key. Take a 128-bit password system as an example. It has 2nS types of key combinations. Suppose the wrongdoers use a computer that can verify 40,000,00,00,000 keys per second to guess the decryption key. 'They still have to spend 2 丨 28 / (4χ 1〇8 × 6〇 × 60〇24 × 365 ) # 2 7 × 1022 until ^ Paper New Zealand National Standard (CNS) (210X297mm) --- Please read the notice on the back before filling out this page. 5 4 〇4 A7 — 一 _ _B7______ 5. Description of the invention (13) The inspection is completed. Such a time is too great, so our voice encryption system can completely resist the attack of the ciphertext attack method. 2. Known plaintext attack Known-Plaintext Attack: This method assumes that the attacker has some secret speech combination before and after decryption. In this case, the attacker may obtain the decryption gold by analyzing the secret speech before and after decryption. .If the sender still uses the same key to encrypt the secret voice next time, then the attacker can use this key to correctly extract his secret voice. In order to prevent this attack method, we use a different gold each time Key to encrypt the secret voice, so that even if one of the secret keys is captured by the criminal, they still cannot use this secret key to decrypt the next secret voice, so in this case, our encryption system can resist Attack by plaintext attack. ---------- Seed coat --- „--- Γίΐτ ------ ^ (Please read the precautions on the back / fill in this page first) Central Standard of the Ministry of Economy The paper size of the seal I · 1 sheet of the Bureau Cooperative Consumer Cooperative is applicable to the Chinese National Standard (CNS) A1 2 3 4 specification (21〇 × 295 mm) 15 1. Chosen-Text Attack: This method assumes an attack They not only have some secret speech combinations before and after decryption, but also have the ability to select or control this secret speech. Like the plaintext attack method, the attacker may also obtain the decryption key by analyzing the secret voice before and after decryption; but as long as the encrypted key is different each time, our system can resist this. Attack 453〇46 453〇46 Staff Consumer Cooperative Indo A7 of the Central Standards Bureau of the Ministry of Economic Affairs A ~~ 'B7 V. Description of the invention (14) " 一 < Schematic description > Party encryption, schematic. FIG. 2 is a schematic diagram of the decryption structure of the touch side of the voice of the present invention. Fig. 3: Schematic diagram of deduction of confidential voice of the sender according to the present invention. Figure 4. A schematic diagram of the general voice processing flow of the present invention on the transmitting side. Fig. 5 'shows the general voice disambiguation and concealment in the receiving party # | taken the flow diagram. Fig. 6 ·· this picture is in the receiver's _voicena flow diagram. Fig. 7 'A schematic diagram of a waveform of an original voice used in the experiment of the present invention. Figure VIII. Schematic waveform of the disguised speech obtained by the present invention in an experiment. Fig. 9: Schematic diagram of error waveforms of general speech and disguised speech in Experiment 1 of the present invention. FIG. 10 is a waveform diagram of the original confidential voice used in the experiment 1 of the present invention. Figure 11: Schematic diagram of confidential speech waveforms taken from disguised speech in Experiment 1 of the invention 0 Figure 11: Schematic error waveforms of confidential speech processed by the invention in Experiment 1 before and after processing 0 Figure 13: The present invention in Experiment 2 The waveform diagram of the original general speech used in. Fig. 14: Schematic waveform of disguised speech obtained by the present invention in Experiment 2. Figure 15: Schematic diagram of error waveforms of general speech and camouflage speech in Experiment 2 of the present invention. The paper size is suitable for financial standards (cns) from specifications (training 297). (Please read the precautions on the back first. Fill out this Page), installed--i-m 16