A7A7
五、發明說明（ 【技術領域】 本發明係、關於—種語音訊號f景雜訊抑制系統及方 > ’主要是針對語音短時間與糾間㈣設計之背景 雜§fl抑制係統及方法。【先前技術】 語音訊號是通訊系統中所傳送最主要的資料型能，在 通訊的過程中，除了語音訊號之外，通話環境中的背景雜 訊也會伴隨著進入電話中，因而造成若干程度的干擾並進 而影響通話品質；尤其是最近快速成長的行動通訊電話， f是容㈣受背景雜訊的影響，因此抑制背景雜訊的技術 疋在當則強調服務品質的通訊系統中一個重要課題，常用 的抑制背景雜訊技術有下列三種： 第一種方法是頻率領域雜訊消去法。這種方法的基 本㈣是在非語音的區段，估計此時雜訊在頻率領域上= 2量，級在接下來的語音區段巾，於解領域中減掉之 :所估計在各個頻率上的雜訊能量。這種方法雖然簡 単，但-般的背景雜訊的統計特性會隨時間而改變，因此 其抑制背景雜訊的效果有限。在美國專利挪議2和 US05=2927中，有提到使用頻率領域雜訊消去法的概念。 第二種方法是時間領域雜訊消去法。這種方法 原理是利用兩支麥克風來接收外部訊號，第一支麥克^主 要是接收語音以及背景雜訊的訊號，第收:景雜訊的訊號。因此，藉由第二支麥克風就 丨出月訊的大小’然後在時間領域上將第一支麥克風的 U 張尺.财 10 15 20 (請先閱讀背面之注意事項再填寫本頁) « ---------線一____一___.__, —I—I—I—IIIII ✓ I I I I . 526468 A7V. Description of the Invention (Technical Field) The present invention relates to a system and method for suppressing voice signal f scene noise suppression > 'It is mainly a background noise suppression system and method designed for speech short time and time correction. [Previous technology] Voice signals are the most important data type capability transmitted in communication systems. In addition to voice signals during communication, background noise in the call environment will also accompany the phone, which will cause a certain degree. Interference, which in turn affects call quality; especially recently the rapid growth of mobile communication phones, f is tolerated by background noise, so the technology to suppress background noise is an important issue in communication systems that emphasize service quality There are three commonly used background noise suppression techniques: The first method is the noise cancellation method in the frequency domain. The basic method of this method is in the non-speech section. It is estimated that the noise in the frequency domain at this time = 2 The level in the next speech segment is subtracted from the solution field: the estimated noise energy at each frequency. This method is simple, but- The statistical characteristics of general background noise will change over time, so its effect of suppressing background noise is limited. In US Patent No. 2 and US05 = 2927, the concept of using the frequency domain noise cancellation method is mentioned. The two methods are the time-domain noise cancellation method. The principle of this method is to use two microphones to receive external signals. The first microphone ^ is mainly used to receive signals of voice and background noise, and the first one is the signal of scene noise. Therefore, with the second microphone, the size of the monthly news is displayed, and then the U ruler of the first microphone is used in the time field. Cai 10 15 20 (Please read the precautions on the back before filling this page) «- -------- Line One ____ 一 ___.__, —I—I—I—IIIII ✓ IIII. 526468 A7
526468 A7 經濟部智慧財產局員工消費合作社印製 五、發明說明（> ) 的品質，而且可以適應性地調整相關係數。 本發明之另一目的係在於提供一種語音訊號背景雜訊 抑制系統及方法，其複雜度低，且只需要一支麥克風， 所以相當適合應用在最近快速成長的行動電話以及語音辨 5 識技術中，藉此提高語音編碼品質以及語音辨識率。 【技術内容】 語音訊號背景雜訊抑制系統及方法係用來改善因背景 雜訊的影響所造成語音品質的下降，類比的語音訊號首先 經過取樣器做類比數位訊號之間的轉換，以作為後續的數 10 位訊號處理之用。語音訊號的頻寬是4KHz左右’根據 Nyquist取樣定理，一般的取樣頻率是8KHz，為了提高取樣 訊號之間的相關性，我們將取樣頻率提高四倍’也就是 32KHz的取樣頻率，稱之為超取樣。取樣之後的數位訊 號，我們用12個位元的脈衝編碼調變技術（Pulse Code 15 Modulation，PCM)來表示，也就是說數位語音取樣容許的變 化範圍在± 2048之間。 本發明之語音訊號背景雜訊抑制系統及方法，包括 有··一個超取樣單元，兩個低通渡波器單元，一個適應性 語音分析器單元，一個語音週期檢測器單元，一個背景雜 20 訊抑制濾波器單元，和一個高頻強化器單元。假設包含 背景雜訊的語音訊號為叉(〇 ;首先乂(〇會先由超取樣單元 對它做超取樣（大於兩倍語音頻寬的取樣頻率），並且將取 樣所得的數位訊號，&⑷，利用12位元的脈波編碼調變加 以表示，其中A代表第々個取樣訊號。由於超取樣的關 ______-5- _ 本紙張尺度適用中國國家標準（CNS)A4規格（21G X 297公髮） ' (請先閱讀背面之注意事項再填寫本頁) -裝 · —線— 526468 A7 五 、發明說明（仏） 經濟部智慧財產局員工消費合作社印製 係，語音訊號頻寬以外的雜訊也會被包含進來；因此，在 超取樣單元之後，必須經過一個低通濾波器，移除語音訊 號頻寬以外不必要的訊號。經過第一個低通濾波器的數位 訊號，I⑷，分別進入適應性語音分析器單元、語音週 期檢測器單70、以及背景雜訊抑制濾波器單元來進行下一 步的處理。在適應性語音分析器單元中，利用階數為"的 全極點適應性攄波器來估計語音訊號，全 器的係數是_，一}代表第刺波器係數= 個==出來代表語音訊號獨有特性的滤波器係數，會被 π)达到月，'雜訊抑制遽波器單元；另一方面，㈣會被送 到語音週期檢測器單元去估計語音訊號的週期，估計所得 的週期Ρ範圍在3〜10w,左右，如果取樣頻率是職，則一 個週期所相對應的取樣數目是96〜32〇左右。每個語音訊號 取樣的週期都會被估計，並且也和全極點適應性遽波器的 15係數-樣送到背景雜訊抑制渡波器單元，以進行下一步的 抑制背景雜訊。 立八ί!!雜訊抑制渡波器單元中’利用分別在適應性語 :和語音，期檢測器單元估計所得滤波 器係數 , 曰k / ρ 5又计背景雜訊抑制濾波器。經過第一 20個低通據波器的㈣此時便送進我們所設計的背景雜 2 ί :以降低摻雜在語音訊號中的背景雜訊能量並 比。由於原來語音訊號中的高頻成分也會被 二=鳴波器所抑制，所以我們又設計了一個高頻 __吾音訊號高頻被抑制的成分。最後，再經 本紙張尺度適用中國國家標? 5 線 訊 526468 A7 10 15 經濟部智慧財產局員工消費合作社印製 發明說明（^) 過一個低通濾波器，濾掉語音訊號頻寬以外的雜訊，得到 品質提升的語音訊號，文⑷。 【圖式簡單說明】 。月芩閱以下有關本發明一較佳實施例之詳細說明及其 附圖^可進-步瞭解本發明之技術内容及其目的功效； 有關該實施例之附圖為： 圖-為本發明語音訊號背景雜訊抑制系統之架構圖； 圖二為該語音訊號背景雜訊抑制系統 析器的電路方塊圖； 日刀 一圖三為該語音訊號背景雜訊抑㈣'統之適應性預估滅 波為係數估計電路方塊圖； 、 圖四為該語音訊號背景雜訊抑制系統之 器的電路方塊圖；以及 ^双測 圖五為该語音訊號背景雜訊抑 、壸、士 Α 〜、、、先之为景雜訊抑吿丨丨 濾波的電路方塊圖。 刺 【主要部分代表符號】 101超取樣器 102低通濾波器 103適應性語音分析器 104背景雜訊抑制濾波器 105語音週期檢測器 106南頻強化器 107低通濾波器 21正負判別器 (請先閲讀背面之注意事項再填寫本頁) 裝 ir°J. 本紙張尺⑦霸？關家鮮（CNS)A4規格 297公釐） 526468 Α7 __ Β7 五、發明說明（ PA0IG2fle.TWP - 8Π 0 10 經濟部智慧財產局員工消費合作社印製 15 22步階估計器 221適應性步階決定器 23適應性預估濾波器 31正負判別器 41週期檢測器 51雜訊整形濾波器 【較佳實施例】 請參閱圖一所示，本發明之語音訊號背景雜訊抑制系 統’包括有· 一個超取樣器1〇1，兩低通濾波器1〇2、1〇7， 一個適應性語音分析器1〇3，一個語音週期檢測器1〇5，一 個背景雜訊抑制濾波器104，和一個高頻強化器1〇6。在正 式進行背景雜汛抑制處理之前，會先經過前置處理，將類 比語音訊號轉換成適合後續處理的數位訊號，包括超取樣 以及低通濾、波，超取樣器101 一方面對外部的類比語音訊 號做類比數位轉換’另一方面將轉換的數位訊號用脈波編 碼調變(Pulse Code ModulationJPCM)方式加以表示。在進行類 比數位轉換時，取樣頻率遠大於取樣定理所規定的最低頻 率，以提咼取樣之間的相關性。在本實施例中建議的取樣 頻率是32从fe，也就是一般語音頻寬从份的8倍。低通濾波 裔102是用以移除語音頻寬之外的雜訊，尤其是之前經過 超取樣器101，更需要利用一個低通濾波器1〇2將訊號的頻 寬限制.在語音的頻寬之内，以提高後續處理單元的效能。 在本貫施例中’採用一個三階Butterworth的低通濾波器， 截止頻率設定在語音的頻寬4紛·ζ。低通濾波之後的訊號 (請先閱讀背面之注意事項再填寫本頁) «. 訂： i線- (cns)a4 m^10 x 297^t)— ---~ ____ 526468 經濟部智慧財產局員工消費合作社印製 五 、發明說明（2 ) "PAO ICUA^'.TWP - y/ la l(A〇刀別進入適應性語音分析器1〇3、語音週期檢測器 105、以及月厅、雜訊抑制濾波器1〇6，以進行下一階段的處 理 圖一疋適應性語音分析器電路方塊圖。適應性語音分 析為103包含—個正負判別器21、-個步階估計器22、以 及一個適應性預估濾波器23。正負判別H21根據輸入語音 訊號I㈨及適應性語音分析器103所估計的&㈨的大小， 來決定輸出位元6(A)的值。 b(k) ^fSnn(k)>Se(k) ^^nn(^) < Se(k) ⑴ 步階估计為22是利用之前所決定的位元來估計現在取 樣的步階，這個步階就代表適應性預估濾波器23的預估殘 餘之補償。假設目前決定出來的位元是咐），則在步階估 計器22内的適應性步階決定器會根據咐)及其前三個位元 咐-1)、蚴—2)、蚴—习來判定目前適應性語音分析器幻的 狀態，並決定出-個修正的係數，咐，如表一所示。之 後，再利用-個-階的迴授平均器產生在時間^所估計 的步階幻，如下所示：5{k) - 5(k + a(k) (2) ，其中/?<1是迴授平均線路的常數，用以控制平均的 長度。P越接近1則平均的長度越長，但反應的時間越 慢，一般的大小是〇.9左右飞是用以調整修正係數邮)大 小的常數，以使適應性語音分析器1〇3可以跟隨語音訊號 的變化。最後，γ階適應性預估濾波器23利用過去"個語 1 本紙張尺度適用中_ ^準（CNS)A4規格（21Q χ 297公[ 10 15 20 (請先閱讀背面之注意事項再填寫本頁) |線· 526468 A7 五、發明說明（ nQ2QQ--TWP - 10/10 10 15* 經濟部智慧財產局員工消費合作社印製 20 音估计值以及步階估計值Μ幻，產生下一個語音取樣的估 計值以hi) ^ N-\ ^k^)^ai(krsu(k-i)+s(k) (3) 表一是適應性步階決定器221的參考表。修正的係數 α(々)就是根據這張表所決定出來的。如果連續四個位元都 疋一樣的話，表示適應性語音分析器1〇3所估計的&⑷不 夠大’所以將修正係數 <幻設為2，使適應性語音分析器 103可以快速追上語音訊號的變化。如果只是其中連續三 個位兀都是一樣的話，則給一個稍小的修正係數α(幻=1， 以稍微增加步階的大小。如果連續四個位元都不一樣的 活，則將修正係數設為_丨。因為此時適應性語音分析器 103咼估了語音訊號，所以必須修正回來。在其餘的情況 下 <幻=0 ’代表適應性語音分析器103追的到語音訊號的 變化。 圖三是適應性預估濾波器23係數估計電路方塊圖，用 以產生W階適應性預估濾波器23的#個係數， %⑷，/ = 1，2,··.，#。適應性預估濾波器23係數估計包含一個正 負判別器31、兩排長度為#_丨分接式延遲線、一排長度為 iV的一階的迴授平均器。兩個輸入訊號，包括語音訊號估 計值5；㈨以及位元值。語音訊號估計值乂⑷在進入上 面一排分接式延遲線前，會先經過正負判別器31，根據其 正負號而分給予+ 1、一；[的值。這樣可以降低計算量。位 元值0(幻在進入下面一排分接式延遲線前，會先乘上一個 μ氏張尺心用?(CNS)A4ii^ -10- χ 297^1) (淨先閱讀背面之注意事項再填寫本頁) __ -----^------------------------------------ 526468526468 A7 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs 5. The quality of the invention description (>), and the correlation coefficient can be adjusted adaptively. Another object of the present invention is to provide a system and method for suppressing background noise of a voice signal, which has low complexity and requires only a microphone, so it is quite suitable for application in recent fast-growing mobile phones and speech recognition technology. To improve speech encoding quality and speech recognition rate. [Technical content] The background noise suppression system and method for speech signals are used to improve the degradation of speech quality caused by the background noise. The analog speech signals are first converted by the sampler to analog digital signals for subsequent use. For digital signal processing. The bandwidth of the voice signal is about 4KHz. According to the Nyquist sampling theorem, the general sampling frequency is 8KHz. In order to improve the correlation between the sampling signals, we increase the sampling frequency by four times, that is, the sampling frequency of 32KHz, which is called super sampling. The digital signal after sampling is expressed by 12-bit Pulse Code 15 Modulation (PCM). That is to say, the allowable change range of digital speech sampling is within ± 2048. The background noise suppression system and method for speech signals of the present invention include: an oversampling unit, two low-pass wave waver units, an adaptive speech analyzer unit, a speech period detector unit, and a background noise 20 signal Suppression filter unit, and a high-frequency booster unit. Assume that the voice signal containing background noise is a cross (0; first 乂 (0 will be oversampled by the oversampling unit (more than twice the sampling frequency of speech and audio width), and the sampled digital signal will be & ⑷, expressed by 12-bit pulse wave code modulation, where A represents the first sampling signal. Because of the oversampling ______- 5- _ This paper size is applicable to China National Standard (CNS) A4 specification (21G X 297)) (Please read the precautions on the back before filling out this page) -Installation · —Line — 526468 A7 V. Description of Invention (仏) Printing Department of Employees' Cooperatives, Intellectual Property Bureau, Ministry of Economic Affairs, outside of voice signal bandwidth Noise will also be included; therefore, after the oversampling unit, a low-pass filter must be passed to remove unnecessary signals outside the bandwidth of the voice signal. The digital signal after the first low-pass filter, I⑷ , Into the adaptive speech analyzer unit, speech cycle detector unit 70, and background noise suppression filter unit for the next step of processing. In the adaptive speech analyzer unit, An all-pole adaptive wave filter of order " is used to estimate the speech signal, and the coefficient of the whole device is _, a} represents the first waverbinator coefficient = === the filter coefficients representing the unique characteristics of the voice signal, will Π) reaches the month, 'the noise suppresses the waver unit; on the other hand, it will be sent to the speech period detector unit to estimate the period of the speech signal, and the estimated period P ranges from 3 to 10w, if The sampling frequency is a duty, then the number of samples corresponding to a cycle is about 96 ~ 32. The sampling cycle of each voice signal will be estimated, and it will be sent to the background with the 15 coefficients of the all-pole adaptive wavelet. The noise suppression wave filter unit is used to suppress the background noise in the next step. Li Ba ί !! The noise suppression wave filter unit 'uses the adaptive filter: and the speech and period detector units to estimate the obtained filter coefficients, respectively. The k / ρ 5 is also considered as the background noise suppression filter. After passing through the first 20 low-pass data wave filters, the background noise 2 designed by us is now sent to reduce the background doped in the speech signal. Noise energy is not compared. The high-frequency component in the original voice signal will also be suppressed by the two = sounder, so we have designed a high-frequency __ 吾 音 signal high-frequency suppressed component. Finally, according to this paper scale, the Chinese national standard is applied ? 5 line news 526468 A7 10 15 Printed invention description printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs (^) Pass a low-pass filter to filter out noise outside the bandwidth of the voice signal and get a voice signal with improved quality. [Brief description of the drawings]. Please read the following detailed description of a preferred embodiment of the present invention and its accompanying drawings ^ You can further understand the technical content of the present invention and its purpose and effectiveness; the drawings related to this embodiment For: Figure-This is the architecture diagram of the background noise suppression system of the voice signal of the present invention; Figure 2 is a circuit block diagram of the background noise suppression system analyzer of the voice signal; 'The adaptive prediction of the system is a block diagram of the coefficient estimation circuit; Figure 4 is a circuit block diagram of the device for the background noise suppression system of the voice signal; and ^ Double test chart 5 is the voice signal King noise suppression, Kun, Shi Α ~ ,,, view of the first noise suppression Gaoshu Shu filtering circuit block diagram. [Symbols of main parts] 101 Supersampler 102 Low-pass filter 103 Adaptive speech analyzer 104 Background noise suppression filter 105 Speech period detector 106 South frequency enhancer 107 Low-pass filter 21 Positive and negative discriminator (please (Please read the precautions on the back before filling out this page) Loading ir ° J. This paper ruler? Guan Jiaxian (CNS) A4 specification 297 mm) 526468 Α7 __ Β7 V. Description of the invention (PA0IG2fle.TWP-8Π 0 10 Printed by the Intellectual Property Bureau Employee Consumer Cooperative of the Ministry of Economic Affairs 15 22 Step estimator 221 Adaptive step decision Filter 23 adaptive estimation filter 31 positive and negative discriminator 41 period detector 51 noise shaping filter [preferred embodiment] Please refer to FIG. 1, the voice signal background noise suppression system of the present invention includes: Supersampler 101, two low-pass filters 102, 107, an adaptive speech analyzer 103, a speech period detector 105, a background noise suppression filter 104, and a High-frequency intensifier 106. Before the background noise suppression process is officially carried out, it will be pre-processed to convert the analog voice signal into a digital signal suitable for subsequent processing, including oversampling and low-pass filtering, wave, and oversampling. On the one hand, the converter 101 performs analog digital conversion on the external analog voice signal. On the other hand, the converted digital signal is expressed by Pulse Code Modulation (JPCM) method. When changing, the sampling frequency is much larger than the minimum frequency specified by the sampling theorem, in order to improve the correlation between the samples. In this embodiment, the sampling frequency is recommended to be 32 times fe, which is 8 times the general voice and audio frequency. The low-pass filter 102 is used to remove noise outside the speech and audio width, especially after passing through the supersampler 101, it is necessary to use a low-pass filter 102 to limit the bandwidth of the signal. Within the bandwidth to improve the performance of the subsequent processing unit. In the present embodiment, 'a third-order Butterworth low-pass filter is used, and the cut-off frequency is set to 4 voice bands of speech. The signal after low-pass filtering (Please read the notes on the back before filling in this page) «. Order: i-line-(cns) a4 m ^ 10 x 297 ^ t) — --- ~ 526468 Printed by the Consumers’ Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs 、 Explanation of invention (2) " PAO ICUA ^ '. TWP-y / la l (A〇 knife enters adaptive speech analyzer 103, speech period detector 105, moon hall, noise suppression filter 1 〇6, for the next stage of processing-Adaptive speech analyzer circuit Block diagram. The adaptive speech analysis 103 includes a positive-negative discriminator 21, a step estimator 22, and an adaptive estimation filter 23. The positive-negative discrimination H21 is based on the input speech signal I㈨ and the adaptive speech analyzer 103. The estimated size of & ㈨ determines the value of output bit 6 (A). B (k) ^ fSnn (k) > Se (k) ^^ nn (^) < Se (k) 步 step The order estimation of 22 is to use the previously determined bits to estimate the current sampling step. This step represents the compensation of the estimated residual of the adaptive prediction filter 23. Assuming the currently determined bits are commanded), the adaptive step determiner in step estimator 22 will be based on commanded) and its first three bits commanded -1), 蚴 -2), 蚴-Xi To determine the current state of the adaptive speech analyzer, and determine a modified coefficient, as shown in Table 1. After that, the feedback averager of one order is used to generate the step imagination estimated at time ^ as follows: 5 {k)-5 (k + a (k) (2), where /? ≪ 1 is the constant of the feedback average line, used to control the average length. The closer P is to 1, the longer the average length, but the slower the reaction time, the general size is about 0.9. Fly is used to adjust the correction factor. ), So that the adaptive speech analyzer 103 can follow the change of the speech signal. Finally, the γ-order adaptive estimation filter 23 uses the past " Phrase 1 This paper size is applicable _ ^ Standard (CNS) A4 specifications (21Q χ 297 public [10 15 20 (Please read the precautions on the back before filling (This page) | Line · 526468 A7 V. Description of the invention (nQ2QQ--TWP-10/10 10 15 * The 20-tone estimate and the step estimate MU are printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs to generate the next voice The estimated value of the sample is hi) ^ N- \ ^ k ^) ^ ai (krsu (ki) + s (k) (3) Table 1 is a reference table of the adaptive step determiner 221. The modified coefficient α (々 ) Is determined according to this table. If the four consecutive bits are all the same, it means that & 估计 estimated by the adaptive speech analyzer 10 is not large enough, so the correction coefficient < magic is set to 2 , So that the adaptive speech analyzer 103 can quickly catch up with changes in the speech signal. If only three consecutive bits are the same, give a slightly smaller correction coefficient α (magic = 1, to slightly increase the step Size. If four consecutive bits are not the same, set the correction factor to _ 丨. Because at this time The adaptive speech analyzer 103 estimates the speech signal, so it must be corrected. In the rest of the cases, <Magic = 0 'represents the change in the speech signal tracked by the adaptive speech analyzer 103. Figure 3 shows the adaptive prediction. Filter 23 coefficient estimation circuit block diagram for generating # coefficients of the W-order adaptive estimation filter 23,% ，, / = 1, 2, ...., #. Adaptive estimation filter 23 coefficient estimation Contains a positive and negative discriminator 31, two rows of #_ 丨 tapped delay lines, and a first-order feedback averager of length iV. Two input signals, including the estimated value of the voice signal 5; ㈨ and bits The estimated value of the voice signal: before entering the upper row of tapped delay lines, it will first pass the positive and negative discriminator 31, and will be given +1, one according to its positive and negative sign; []. This can reduce the amount of calculation. Bit value 0 (phantom will be multiplied by a μ's ruler before entering the next row of tapped delay lines? (CNS) A4ii ^ -10- χ 297 ^ 1) (Read the note on the back first (Please fill in this page for matters) __ ----- ^ ------------------------------------ 526468
發明說明（1 小於1常數增益e。V個輸出訊號，也就是適應性預估濾 波器23的〜個係數，軌/=1，2，...^。以第/個係數_為 例（z # 1 )，我們可以用下面的方程式加以表示其產生的 方式： 5 = ^ ~ 1) + e * b(k) * SGN[Se(k)] (4) 其中，d是一個常數，代表一階的迴授平均器所平均 的長度，一般的大小是〇.9左右。SGN[]代表正負判別器的 動作，也就是取中括號内取樣的正負號。基本上，方程式 (4)就是一個簡化的基於隨機梯度最降搜演法。在適應性預 10估濾波器23的λΜ固係數的產生方法中，％㈨的產生方法多 了一個小於1的常數項γ : a,{k) = -1) + e * b{k) * SGN[Se(k)] + / (5、 这是為了表示目前要估計語音訊號和前一個語音訊號 估計值之間強烈的相關性。 15 圖四為語音週期檢測器電路方塊圖，用以估計語音訊 號的週期。語音週期檢測器1〇5包含一排長度為 (κ,η+ι)的分接式延遲線，（Kin+1)個減法器， (Unin +1)個絶對值單元，+1)個一階的迴授平均 器，以及-個週期檢測器41。^代表語音最大可能的週 20期，4代表語音最小可能的週期。如果取樣頻率是 32紐z，則4,32〇，，所以分接式延遲線、減法 器、絕對值單元、以及一階的迴授平均器的長度數目就是 225。語音週期檢測器105的輸入訊號㈣一方面經由上面 的分接式延遲線儲存過去（n+1)個值，另—方面㈣ - 11 本紙張尺錢财關家鮮~~--- (請先閱讀背面之注意事項再填寫本頁) · i線- 經濟部智慧財產局員工消費合作社印製 526468 A7Description of the invention (1 is less than 1 constant gain e. V output signals, that is, ~ coefficients of the adaptive estimation filter 23, orbit / = 1, 2, ... ^. Take the / coefficient _ as an example ( z # 1), we can use the following equation to show how it is generated: 5 = ^ ~ 1) + e * b (k) * SGN [Se (k)] (4) where d is a constant, representing The average length of the first-order feedback averager is generally about 0.9. SGN [] represents the action of the positive and negative discriminator, that is, the positive and negative signs sampled in square brackets. Basically, equation (4) is a simplified search method based on the stochastic gradient minimization. In the method of generating the λM solid coefficient of the adaptive pre-estimation filter 23, the method of generating% ㈨ has a constant term γ less than 1: γ: a, (k) = -1) + e * b (k) * SGN [Se (k)] + / (5. This is to indicate that there is a strong correlation between the currently estimated speech signal and the previous estimate of the speech signal. 15 Figure 4 is a block diagram of the speech period detector circuit, used to estimate The period of the speech signal. The speech period detector 105 includes a row of tapped delay lines of length (κ, η + ι), (Kin + 1) subtracters, (Unin +1) absolute value units, +1) a first-order feedback averager, and a period detector 41. ^ Represents the maximum possible period of speech 20 periods, 4 represents the minimum possible period of speech. If the sampling frequency is 32 NZ, then 4,32, so the number of lengths of the tapped delay line, subtractor, absolute value unit, and first-order feedback averager is 225. The input signal of the speech period detector 105 ㈣ stores the past (n + 1) values through the tapped delay line on the one hand, and-on the other hand ㈣-11 paper rule money and wealth ~~ --- (please (Please read the notes on the back before filling this page) · i-line-Printed by the Intellectual Property Bureau of the Ministry of Economic Affairs, Consumer Cooperatives 526468 A7
tjt去的值相減並取絕對值後，便分別進入-階的迴 ,=上述的動作，就是尋找入⑷和過去語音估計 ▲〗々相關丨生。假設1⑷和過去第户個估計值相關性最 了則第P個延遲所對應的一階的迴授平均器輸出值會最 5 、、所以在週期檢測器41中，便會根據方程式(6)、（7)，檢 測出我們所要的語音週期p，如下所示： " ，tf —4^m，/)|])> 忌 ⑺ 、其中，E[]代表一 P白匕迴授平均器的動作，』[{—()} 10代表求出使括號内的值最小的參數ζ·飞是一階的迴授平 均器輸出值的臨界值，用以區別母音與非母音的一個經驗 值。如果目前的取樣不屬於語音訊號中的母音，則所檢測 的？ = 〇 15 經濟部智慧財產局員工消費合作社印製 20 (¾先閱讀背面之注意事項再填寫本頁) 圖五為背景雜訊抑制濾波器電路方塊圖，用以結合在 語音分㈣和語音„檢_所分卿到的語音特 性係數你)以及語音週期户，進行背景雜訊抑制。背景雜 訊抑制濾波H1G4包含兩排長度為#的分接式延遲線、一 個，遲量為㈣延遲器、…個加法器、_個雜訊整形遽 波為5卜輸人_為語音訊號㈣、語音特性練⑽）、 以及語音週期P，輸出訊號為背景雜訊被 語 號’也就是Snn(k-l)，Snn(k-2)，'s^k-N);下面一排的分 接式延遲線根據語音週期檢測器所檢測到的週期p。先將 - _ 12 _ 本紙張尺度適用中國國家標準（CNS)A4規297公爱了 526468 Α7 Β7 PAO 1028δ.Τν\/Μ ^ Ι5/Ιβ 五、發明說明（丨f) 輸入訊號為語音訊號入㈨延遲p個取樣，然後儲存距離 心(幻/>個取樣之前以及其之前V個語音訊號，也就是 (請先閱讀背面之注意事項再填寫本頁) Λ〇。之後，㈣人㈣，····.·几认—〜和 HPXSJk-P—V)，·····.A#-P—TV)這將兩組訊號依序相加，然 5後和語音特性係數4幻一起送入雜訊整形濾波器51。由於 這兩組吼號内的語音訊號的相似性相當高，所以對語音而 言是同調相加；另一方面，背景雜訊不具這種相似性，所 以是非同調相加。因此，可以達到同調消除雜訊的功效 (如果p = 〇的話，則這一項功能沒有效果）。在雜訊整形濾 10波為51中，將這iV +1個同調相加的語音訊號依下列的整形 濾波器的轉換函數組合起來： 1 - H(z) =—弓-- 1 - ^ α,α/ζ'/ / = 1 •線· 其中’ α和々疋兩個常數，0 ^ Μ α ^ i，用以控制語音 頻譜的波峰波谷的大小。α越接近丨且々越接近0，則波峰 經濟部智慧財產局員工消費合作社印製 15越大波谷越深，但高頻訊號也會被衰減的更多。建議的選 取值疋α=0.9，々=0·2。由於α i代表語音訊號的特性，所以經過 整形濾波器51的轉換函數之後，原來的訊號的頻譜會被整 形成類似語音訊號的形狀，也就是說背景雜訊的頻譜會跟 隨語音訊號的頻譜而變化，因此就產生了所謂的遮蔽效應 2〇 (MaskingEffect)，而達到抑制背景雜訊的效果。由於之前 我們先做過同調相加的動作，因此大幅提高遮蔽效應效 —-----— - 13 _ 本紙張尺錢®目家鮮（CNS)A4驗(21^7^-— ___ 526468 A7 五、發明說明（（么) jzo0.TV\ih - 14/Ιΰ 杲 經濟部智慧財產局員工消費合作社印製 接下來’將背景雜訊抑制濾波器1〇4處理後的語音訊 號再送入高頻強化器106。 ^/(ζ) = \-γζ~χ (9) 基本上，這就是一個高通濾波器，〇<?/<1，用以補償 口為正形;慮波為所造成咼頻农減的影響。最後，再經過和 前面一樣的低通濾波器107，以移除因為在適應性背景雜 訊抑制系統中所產生聲音頻寬以外的雜訊。 【特點及功效】 本發明所提供之語音訊號背景雜訊抑制系統及方法， 與岫述引證案及其他習用技術相互比較時，更具有下列之 優點： ^ 、L本發明提供一種語音訊號背景雜訊抑制系統及方 法，一方面利用一個全極點線性預估濾波器來重建語音訊 唬的，i $彳面也檢測出只存在於語音訊號中的週 期。取後，根據估計出來的語音訊號相關性係數以及語音 的週期’將背景雜訊加以抑制，進而提升語音訊號的 品質。 、2 •本發明提供—種語音訊號背景雜訊抑制系統及方 法，可以大幅提昇低訊號雜訊比之輸入訊號的品質，而且 可以適應性地調整相關係數。 、 3.本發明提供一種語音訊號背景雜訊 法，其複雜度低，且只需I一 φ ^ n H先及方 要支夕克風，所以相當適合 μ取成長的行動電話以及語音辨識技術中，藉 _張尺^ t _家鮮 10 15 20 (請先閱讀背面之注意事項再填寫本頁) -線- Γ I —r . 526468 A7 ----—---------B7 _ 玉、發明說明（（，） " " 州_0•下 此提南語音編碼品質以及語音辨識率。 明t:::細說明係針對本發明之-可行實施例之具體說 ’惟㈣施例並非用以限制本發明之專利範圍，凡未脫 離本發明技藝精神所為之等效實施或變更，均應包含於本 5案之專利範圍中。 綜上所述，本案不但在技術思想上讀屬創新，並能較 習用物品增進上述多項功效，應已充分符合新穎性及進步 性之法定發明專利要件，爰依法提出申請，懇請貴局核 准本件發明專利申請案，以勵發明，至便。 (請先閱讀背面之注意事項再填寫本頁) 訂： 丨線- 適 度 尺 張 紙 _J本 經濟部智慧財產局員工消費合作社印製 準 標 家 國 國 A S)After the value of tjt is subtracted and the absolute value is taken, they respectively enter the -step back, == The above action is to find the correlation between the input and past speech estimates ▲〗 々. Assuming that 1⑷ has the most correlation with the first estimated value in the past, the output value of the first-order feedback averager corresponding to the P-th delay will be at most 5, so in the period detector 41, it will be based on equation (6) (7), detecting the desired speech period p, as follows: ", tf —4 ^ m, /) |]) > Danger, where E [] represents the average of a P white dagger feedback "{{(()}" 10 means to find the parameter that minimizes the value in parentheses. Fei is the critical value of the first-order feedback averager output value, an experience to distinguish vowels from non-vowels. value. If the current sample does not belong to the vowel in the voice signal, what is detected? = 〇15 Printed by the Intellectual Property Bureau's Consumer Cooperatives of the Ministry of Economic Affairs 20 (¾Read the precautions on the back before filling out this page) Figure 5 is a block diagram of the background noise suppression filter circuit, which is used to combine speech analysis and speech. Detect the speech characteristic coefficient you got) and the period of the speech, to perform background noise suppression. The background noise suppression filter H1G4 contains two rows of tapped delay lines with a length of #, one, and the delay is a chirp delay. , ... adders, _noise shaping wave is 5 input__ for voice signal, voice characteristics training), and voice period P, the output signal is the background noise signal slogan, which is Snn (kl ), Snn (k-2), 's ^ kN); The tapped delay line in the next row is based on the period p detected by the speech period detector. First, -_ 12 _ This paper standard applies Chinese national standards ( CNS) A4 regulation 297 publicly loved 526468 Α7 Β7 PAO 1028δ.Τν \ / Μ ^ 5 / Ιβ 5. Description of the invention (丨 f) The input signal is the voice signal, and the delay is p samples, and then the distance from the heart (Magic / > Before the sampling and V voice signals before it, that is (please Read the notes on the reverse side and fill in this page) Λ〇. After that, ㈣ 人 ㈣, ......... and HPXSJk-P-V), ... A.-P-TV) this Add the two groups of signals in order, and then send them to the noise shaping filter 51 together with the speech characteristic coefficient 4. Since the similarity of the speech signals in these two groups of roars is quite high, it is Homological addition; on the other hand, background noise does not have this similarity, so it is non-homogeneous addition. Therefore, the effect of homophonic noise removal can be achieved (if p = 〇, this feature has no effect). The 10 waves of noise shaping filter are 51, and the iV +1 coherently added speech signal is combined according to the following transformation function of the shaping filter: 1-H (z) = — 弓-1-^ α, α / ζ '/ / = 1 • Line · where' α and 常数 two constants, 0 ^ Μ α ^ i, are used to control the size of the peaks and troughs of the speech spectrum. The closer α is to 丨 and 々 is closer to 0, then The peak printed by the Consumer Property Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs of the 15th, the larger the trough, the deeper, but the high-frequency signal will be more attenuated. Suggested choice疋 α = 0.9, 々 = 0 · 2. Because α i represents the characteristics of the voice signal, after the conversion function of the shaping filter 51, the spectrum of the original signal will be shaped into a shape similar to the voice signal, that is, the background The spectrum of noise will change with the spectrum of the voice signal, so the so-called masking effect 20 (MaskingEffect) is generated, and the effect of suppressing background noise is achieved. Since we have done the same operation before, we have significantly increased Improving the effect of shadowing effect —-----—-13 _ This paper ruler ® mesh home fresh (CNS) A4 test (21 ^ 7 ^-— ___ 526468 A7 V. Description of the invention ((?) Jzo0.TV \ ih -14 / Ιΰ 印 Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs, the next step is to send the voice signal processed by the background noise suppression filter 104 to the high-frequency enhancer 106. ^ / (ζ) = \ -γζ ~ χ (9) Basically, this is a high-pass filter, 〇 <? / < 1, which is used to compensate for the mouth shape; consider the wave to reduce the frequency of agricultural noise reduction Impact. Finally, the same low-pass filter 107 as before is passed to remove noise other than the sound and audio width generated in the adaptive background noise suppression system. [Features and effects] The system and method for suppressing background noise of speech signals provided by the present invention have the following advantages when compared with the cited citations and other conventional technologies: ^, L The present invention provides a background noise background of speech signals The signal suppression system and method, on the one hand, use an all-pole linear prediction filter to reconstruct the speech signal, and the i $ 彳 plane also detects the period that only exists in the speech signal. After that, the background noise is suppressed based on the estimated correlation coefficient of the speech signal and the period of the speech, thereby improving the quality of the speech signal. 2 • The present invention provides a background noise suppression system and method for voice signals, which can greatly improve the quality of input signals with low signal-to-noise ratios, and can adaptively adjust correlation coefficients. 3. The present invention provides a background noise method for voice signals, which has low complexity and requires only I-φ ^ n H to be supported first, so it is quite suitable for μ-growth mobile phones and speech recognition technology. In the middle, borrow _ 张 尺 ^ t _ 家 鲜 10 15 20 (Please read the notes on the back before filling this page) -line- Γ I —r. 526468 A7 ---------------- -B7 _ Jade, invention description ((,) " " State_0 • The following mentions Nanan's speech encoding quality and speech recognition rate. Ming t ::: Detailed description is specific to the -feasible embodiment of the present invention 'However, the examples are not intended to limit the patent scope of the present invention. Any equivalent implementation or change that does not depart from the technical spirit of the present invention should be included in the patent scope of this 5 case. In summary, this case is not only in Technical thinking is an innovation, and can improve the above-mentioned multiple effects compared with conventional items. It should have fully met the statutory invention patent requirements of novelty and progress, apply according to law, and ask your office to approve this invention patent application to encourage invention. , Please. (Please read the notes on the back before filling (This page) Order: 丨 Line-Moderate Rule Paper _J this printed by the Consumers' Cooperative of Intellectual Property Bureau of the Ministry of Economic Affairs
526468 M'濟部智慧財產局員工消費合作社印製 A7 B7 -—-PA010388.TWP 16/18 五、發明說明（（f) 表一：適應性步階決定器參考表 b(n) b(n-l) b(n-2) b(n-3) a(n) -1 -1 -1 -1 2 1 1 1 1 2 -1 -1 -1 1 1 1 1 1 -1 1 -1 1 1 1 1 1 -1 -1 -1 1 1 1 -1 -1 0 -1 -1 1 1 0 -1 1 1 -1 0 1 -1 -1 1 0 -1 -1 1 1 0 1 1 -1 _1 0 1 -1 1 1 0 -1 1 -1 -1 0 -1 1 _1 1 -1 1 -1 1 -1 -16- (請先閱讀背面·之注意事項再填寫本頁) 夤 訂----- 線 74: 本紙張尺度適用中國國家標準（CNS)A4規格（210 X 297公爱）526468 M 'printed by the Consumers ’Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs A7 B7 ----- PA010388.TWP 16/18 V. Description of the invention ((f) Table 1: Reference table for adaptive step determiner b (n) b (nl ) b (n-2) b (n-3) a (n) -1 -1 -1 -1 2 1 1 1 1 2 -1 -1 -1 1 1 1 1 1 -1 1 -1 1 1 1 1 1 -1 -1 -1 1 1 1 -1 -1 0 -1 -1 1 1 0 -1 1 1 -1 0 1 -1 -1 1 0 -1 -1 1 1 0 1 1 -1 _1 0 1 -1 1 1 0 -1 1 -1 -1 0 -1 1 _1 1 -1 1 -1 1 -1 -16- (Please read the precautions on the back and then fill out this page) Order ---- -Line 74: This paper size applies to China National Standard (CNS) A4 (210 X 297 public love)