【正文】
k),S(λ, k)和N(λ, k)分別代表帶噪語音、純凈語音和噪聲的短時(shí)傅立葉變換后的幅度, 和分別代表語音缺失和語音存在概率假設(shè)。 為了達(dá)到上式的噪聲估計(jì),首先對(duì)輸入的每一幀信號(hào)用下式進(jìn)行頻域平滑: (427)其中,b(i)為加權(quán)系數(shù),Y(λ, k)是第λ 幀的第k 個(gè)頻率點(diǎn)的幅度,2w +1是進(jìn)行頻域平滑的相鄰頻點(diǎn)數(shù)。 計(jì)算局部能量最小值局部能量最小值等于前一幀局部能量最小值與當(dāng)前幀帶噪信號(hào)功率譜中較小的一個(gè),即:9 (429)同時(shí)定義一個(gè)臨時(shí)變量,表示如下: (430)每當(dāng)處理第nL幀時(shí),更新局部能量最小值和臨時(shí)變量,分別為: (431)其中,L為一個(gè)包含多幀的窗口大小,決定了局部最小值搜索的范圍,考慮到語音的連續(xù)性和噪聲的改變。之所以這樣判斷是因?yàn)椋涸谡Z音盲點(diǎn)時(shí)刻,帶噪語音功率譜接近于其局部最小值,因此,在上式中比率越小,處于語音間隙的可能性就越大。δ 取值5,I (λ , k)是上式中的指標(biāo)函數(shù)。如下式: (435)其中, 一種改進(jìn)的最小統(tǒng)計(jì)量控制遞歸平均噪聲估計(jì)算法 改進(jìn)的噪聲估計(jì)算法 設(shè)x(n)和d(,1)分別表示干凈語音信號(hào)和加性噪聲信號(hào),則帶噪語音信號(hào)y(n)=x(n)+d(n)。對(duì)帶噪語音信號(hào)分幀并變換到頻域可以得到Y(jié)(k,1)=x(k,Z)+D(k,Z),k為頻帶序號(hào),l為幀序號(hào)。設(shè)為語音無聲段的固定平滑因子,則可以表示為: (437)假定X(k,l)和D(k,l)均滿足復(fù)高斯分布,則在噪聲功率譜給定條件下語音存在概率p(k,l)由下式給出: (438)設(shè)為后驗(yàn)信噪比, 為先驗(yàn)信噪比,可以采用判決反饋的方法獲得,為語音存在時(shí)干凈語音信號(hào)功率譜,為先驗(yàn)語音不存在概率,受語音最小統(tǒng)計(jì)量控制,通過兩次平滑和最小統(tǒng)計(jì)量跟蹤實(shí)現(xiàn)。設(shè)為時(shí)域平滑參數(shù),6(f)為長度為2w+1的歸一化窗,取為W=1的漢寧窗,則平滑過程為: (439) (440)對(duì)平滑輸出信號(hào)s(k,l)進(jìn)行最小值搜索。設(shè) ,和為判決門限,根據(jù)和進(jìn)行語音存在概率硬判決: (443)根據(jù)語音存在概率硬判決去除強(qiáng)語音成分后,進(jìn)行第二次平滑和最小值跟蹤。定義和: (446)設(shè),則先驗(yàn)語音不存在概率為: (447) 實(shí)驗(yàn)仿真本實(shí)驗(yàn)的兩種帶噪語音文件,一種是信噪比為5dB的平穩(wěn)帶噪語音,噪聲為高斯白噪聲。其噪聲跟蹤能力與IMCRA算法相當(dāng)。該算法可以廣泛地應(yīng)用于語音增強(qiáng)系統(tǒng),能夠有效地提高信噪比,抑制音樂噪聲。它廣泛地應(yīng)用于語音通信的背景噪聲抑制、語音壓縮編碼和語音識(shí)別的前端預(yù)處理中。因此,噪聲估計(jì)是語音增強(qiáng)系統(tǒng)中非常重要的一個(gè)部分,估計(jì)的好壞會(huì)直接影響最終的增強(qiáng)效果。本文主要研究了: Martin 提出的基于最小值跟蹤和最小統(tǒng)計(jì)的噪聲估計(jì)算法,Cohen 等提出的最小值遞歸平均算法,并對(duì)這些算法進(jìn)行了仿真實(shí)驗(yàn)和分析。該算法可以廣泛地應(yīng)用于語音增強(qiáng)系統(tǒng),能夠有效地提高信噪比,并且能夠有效地抑制音樂噪聲。 目前存在的問題及今后的發(fā)展方向由于本人所學(xué)有限,本論文提出的改進(jìn)和想法是一些很基礎(chǔ)的,還不全面,需要進(jìn)一步的探索和完善。(2)進(jìn)一步完善噪聲功率譜的估計(jì)算法,對(duì)于許多新型的算法加以研究,進(jìn)一步將噪聲估計(jì)方法和其他方法相結(jié)合,爭取得到更加精確的噪聲估計(jì)。參考文獻(xiàn)[1] 張雄偉,陳亮,楊吉斌.現(xiàn)代語音處理技術(shù)及應(yīng)用[M].北京:機(jī)械工業(yè)出版社,2003:412.[2] 趙立.語音信號(hào)處理[M].北京:機(jī)械工業(yè)出版社,2003:510.[3] R. Martin. Spectral subtraction based on minimum statistics[C], Seventh EuropeanSignal Processing Conference, 1994, 9(5):11821185.[4] 張雪英,張剛,馬建芬,語音處理與編碼[M],北京,兵器工業(yè)出版社,2000.[5] , B. Berdugo. Noise estimation by minima controlled recursive averaging for robust speech enhancement [J], IEEE Signal Proc. Letters, 2002, 9(1):1215.[6] 王華奎,張立毅,數(shù)字信號(hào)處理理論及應(yīng)用[M],太原,太原理工大學(xué)出版社,2005,[7] 姜琳峰,石鴻凌,孫洪,基于最優(yōu)平滑和統(tǒng)計(jì)最小的語音增強(qiáng)[J],武漢大學(xué)學(xué)報(bào)(理學(xué)版),2004,50(1),[8] 陳照平,馬建芬,張雪英,一種基于快速噪聲估計(jì)的MMSE語音增強(qiáng)算法[J],計(jì)算機(jī)工程與應(yīng)用,2007,(22),[9] 張波,曹志剛,低信噪比條件下的一種自適應(yīng)有聲/無聲判決算法[J],信號(hào)處理,1996,12(3) ,[10] 楊行峻,遲惠生等.語音信號(hào)數(shù)字處理[M].北京:電子工業(yè)出版社,1995:1622.[11] 焦人杰,侯麗敏.一種快速自適應(yīng)噪聲譜估計(jì)方法[J].聲學(xué)技術(shù),2007,26(4):735740.[12] 胡航.語音信號(hào)處理[M].哈爾濱:哈爾濱工業(yè)大學(xué)出版社,2002:2029.[13] 易克初,田斌,付強(qiáng).語音信號(hào)處理[M].北京:國防工業(yè)出版社,2000:2230.[14] 韓紀(jì)慶,張磊等.語音信號(hào)處理[M].北京:清華大學(xué)出版社,2004:3041.[15] 韓韜.基于強(qiáng)背景噪聲下的語音端點(diǎn)檢測(cè)算法與實(shí)現(xiàn)[D].長沙:湖南大學(xué),2007.[16] 金學(xué)驥.語音增強(qiáng)算法的研究與實(shí)現(xiàn)[D].杭州:浙江大學(xué),2005.[18] 周為,邱秀清,朱敬鋒等.一種基于噪聲快速跟蹤的語音增強(qiáng)算法[J].電聲技術(shù),2007,31(11):5560.[19] 楊海.感知語音質(zhì)量評(píng)價(jià)PESQ及其在通信系統(tǒng)中的應(yīng)用[J].江西通信科技,2004,2004(2):3647.[20] 朱健華,語音增強(qiáng)方法的研究[D],大連,大連理工大學(xué),2002[21] 胡航.語音信號(hào)處理[M].哈爾濱:哈爾濱工業(yè)大學(xué)出版社,2002:2029.[22] 王永紅,周德新,熊軍.譜減法技術(shù)中“音樂噪聲”的抑制[J].桂林電子工業(yè)學(xué)院學(xué)報(bào),1998,18(1):3842.致 謝時(shí)光流逝,短短四年的大學(xué)生活即將結(jié)束,在論文即將完成之際,感慨頗多。從畢業(yè)論文的選題、研究到撰寫、修改,再到最終的完成,整個(gè)過程賈老師都耐心指導(dǎo),使我能夠從中不斷的學(xué)習(xí)和提高。其次,感謝跟我相處四年的同學(xué)和朋友,在我困難的時(shí)候給予我的鼓勵(lì)和幫助。附錄:外文文獻(xiàn)Noise estimation Algorithms for Speech Enhancement in highly nonstationary EnvironmentsAnuradha R. Fukane1, Shashikant L. Sahare21,2 Electronics and Telemunication departmentCummins college of Engineering For Women, Pune 411052, Maharashtra, IndiaAbstractA noise estimation algorithm plays an important role in speech enhancement. Speech enhancement for automatic speaker recognition system, Man–Machine munication, Voice recognition systems, speech coders, Hearing aids, Video conferencing and many applications are related to speech processing. All these systems are real world systems and input available for these systems is only the noisy speech signal, before applying to these systems we have to remove the noise ponent from noisy speech signal means enhanced speech signal can be applied to these systems. In most speech enhancement algorithms, it is assumed that an estimate of noise spectrum is available. Noise estimate is critical part and it is important for speech enhancement algorithms. If the noise estimate is too low then annoying residual noise will be available and if the noise estimate is too high then speech will get distorted and loss intelligibility. This paper focus on the different approaches of noise estimation. Section I introduction, Section II explains simple approach of Voice activity detector (VAD) for noise estimation, Section III explains different classes of noise estimation algorithms, Section IV explains performance evaluation of noise estimation algorithms, Section V conclusion.Keywords: speech enhancement, Noise, VAD, FFT, Histogram.1. IntroductionSpeech enhancement plays an important role in numerous applications such as hearing aids184。 cell phones184。 reverberation and speech from other speakers. Therefore the degraded speech ponents need to be processed for the enhancement. Speech enhancement algorithms improve the quality and intelligibility of speech by reducing or eliminating the noise ponent from the speech signals. Improving quality and intelligibility of speech signals reduce listener’s fatigue, improve the performance of hearing aids184。 videoconferencing184。 ≈ | X (λ, k)|178。 (2)Where |Y(λ, k)|178。 is the periodogram of clean speed signal and |D(λ, k) |178。 of the noisy speech over a fixed window length. The periodogram |Y(λ, k)|178。 (3)Where α is the smoothing constant. The above recursive equation in recognized as an IIR Low pass filter, provides a smoothed version of periodogram |Y(λ, k)|178。 of the noisy speech frame.2. Compute the smoothing parameter α (λ, k) using equation.3. Compute the smoothed power spectrum P (λ, k) using equation( 3).4. Compute Bias connection factor βmin (λ, k)5. Search for the minimum psd Pmin (λ, k) over a D Frame window. Update the minimum Whenever V (V D) frames are processed6. Compute α update the noise power spectral density (psd) according to equation?d17