魯棒性哼唱特征研究
發(fā)布時間：2018-08-13 09:12
【摘要】：哼唱檢索系統(tǒng),一種基于內容的多媒體檢索系統(tǒng),作為目前的研究熱點,一直存在哼唱特征不穩(wěn)定的問題,造成此問題的主要原因是人聲個性化,哼唱人的音域不同、哼唱節(jié)奏不同、個別音高不準確等問題,都會導致哼唱特征不穩(wěn)定。本文針對哼唱檢索系統(tǒng)中存在的哼唱特征不穩(wěn)定問題,進行了以下幾點研究,旨在提取更具魯棒性的哼唱特征:1、哼唱片段特征提取改進算法針對哼唱音頻,提取哼唱音高特征后,為提高哼唱片段與MIDI的匹配度,改進了哼唱片段的特征提取算法。通過分析人聲哼唱頻率分布范圍,哼唱與MIDI的對齊實驗,對哼唱音高進行規(guī)整操作;結合樂理知識,進行音符切分操作;根據人聲個性化進行半音域轉換,以哼唱人的基準頻率代替原440Hz的統(tǒng)一基準頻率,使得哼唱音符特征與midi的音符特征值達到更好的匹配。以上改進方法通過實驗給出算法的有效性。2、提出了一種基于局部統(tǒng)計的哼唱特征提取算法通過對哼唱音符序列在縱向音域分布和橫向時序變化的局部統(tǒng)計,獲得哼唱旋律的直方圖統(tǒng)計特征。此算法在縱向音域上進行區(qū)間分布投影統(tǒng)計;在橫向上進行音符時序變化模式直方圖分布統(tǒng)計,最終獲得縱向與橫向的聯合直方圖特征,并加入均值、極差、方差特征。最后以4段連續(xù)子序列加整段聯合直方圖特征對音符分布進行描述。此算法不同于傳統(tǒng)的以音高或音符直接作為哼唱特征的表示方法,而是將音符轉化為統(tǒng)計特征的形式,保證了哼唱特征的相對穩(wěn)定,對于不同用戶在哼唱中表現的速度、音域、節(jié)奏等方面的差異有很好的容錯性。最后,通過實驗檢驗了此方法的有效性,實驗數據為5000首MIDI,104首哼唱查詢,應用本文提出的哼唱特征提取算法作為哼唱特征,并采用局部敏感哈希算法(Local Sensitive Hash)作為相似特征匹配算法,得到TOP1準確率為86%,TOP5準確率為92%,與原哼唱識別系統(tǒng)中以音符作為特征的結果進行了比較,實驗結果優(yōu)于原始結果。
[Abstract]:Hem retrieval system, a content-based multimedia retrieval system, has been a hot research topic at present. The main cause of this problem is the personalization of human voice and the different phonetic range of humming person. Hem rhythm is different, individual pitch inaccurate and other problems, will lead to the characteristics of humming instability. In this paper, aiming at the instability of humming feature in humming retrieval system, the following researches are carried out to extract the more robust humming feature: 1, and the improved feature extraction algorithm for humming audio. In order to improve the matching degree between humming segment and MIDI, the feature extraction algorithm of humming segment is improved. By analyzing the range of frequency distribution of human humming, the alignment experiment between humming and MIDI, the tonal pitch of humming is regularized; combining with the knowledge of music theory, the notes are segmented, and the semitone conversion is carried out according to the personality of human voice. The standard frequency of humming person is replaced by the unified reference frequency of the original 440Hz, which makes the feature of humming note and the characteristic value of note of midi better match. The above improved method gives the effectiveness of the algorithm through experiments. 2. A local statistics based humming feature extraction algorithm is proposed. The local statistics of the distribution of humming notes in the longitudinal range and the change of horizontal time series are presented. The histogram statistical features of humming melody are obtained. In this algorithm, interval distribution projection statistics are performed on the longitudinal range, and the histogram distribution statistics of the pattern of note timing change are carried out horizontally. Finally, the joint histogram features of longitudinal and horizontal are obtained, and the features of mean value, range difference and variance are added. Finally, the note distribution is described by four successive subsequences plus the whole segment combined histogram feature. This algorithm is different from the traditional representation method in which pitch or note is directly used as the feature of humming, but it transforms the note into the form of statistical feature, which ensures the relative stability of the feature of humming. Rhythm and other aspects of the difference has a good fault-tolerance. Finally, the validity of this method is verified by experiments. The experimental data is 5000 midi 104s humming query, and the Hem feature extraction algorithm proposed in this paper is used as the humming feature. The local sensitive hashing algorithm (Local Sensitive Hash) is used as the similar feature matching algorithm, and the accuracy of TOP1 is 86 and TOP5 is 922.It is compared with the results of the original humming recognition system with the notes as the feature. The experimental results are better than the original ones.
【學位授予單位】：北京郵電大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TN912.3;TP391.3
本文編號：2180532
[Abstract]:Hem retrieval system, a content-based multimedia retrieval system, has been a hot research topic at present. The main cause of this problem is the personalization of human voice and the different phonetic range of humming person. Hem rhythm is different, individual pitch inaccurate and other problems, will lead to the characteristics of humming instability. In this paper, aiming at the instability of humming feature in humming retrieval system, the following researches are carried out to extract the more robust humming feature: 1, and the improved feature extraction algorithm for humming audio. In order to improve the matching degree between humming segment and MIDI, the feature extraction algorithm of humming segment is improved. By analyzing the range of frequency distribution of human humming, the alignment experiment between humming and MIDI, the tonal pitch of humming is regularized; combining with the knowledge of music theory, the notes are segmented, and the semitone conversion is carried out according to the personality of human voice. The standard frequency of humming person is replaced by the unified reference frequency of the original 440Hz, which makes the feature of humming note and the characteristic value of note of midi better match. The above improved method gives the effectiveness of the algorithm through experiments. 2. A local statistics based humming feature extraction algorithm is proposed. The local statistics of the distribution of humming notes in the longitudinal range and the change of horizontal time series are presented. The histogram statistical features of humming melody are obtained. In this algorithm, interval distribution projection statistics are performed on the longitudinal range, and the histogram distribution statistics of the pattern of note timing change are carried out horizontally. Finally, the joint histogram features of longitudinal and horizontal are obtained, and the features of mean value, range difference and variance are added. Finally, the note distribution is described by four successive subsequences plus the whole segment combined histogram feature. This algorithm is different from the traditional representation method in which pitch or note is directly used as the feature of humming, but it transforms the note into the form of statistical feature, which ensures the relative stability of the feature of humming. Rhythm and other aspects of the difference has a good fault-tolerance. Finally, the validity of this method is verified by experiments. The experimental data is 5000 midi 104s humming query, and the Hem feature extraction algorithm proposed in this paper is used as the humming feature. The local sensitive hashing algorithm (Local Sensitive Hash) is used as the similar feature matching algorithm, and the accuracy of TOP1 is 86 and TOP5 is 922.It is compared with the results of the original humming recognition system with the notes as the feature. The experimental results are better than the original ones.
【學位授予單位】：北京郵電大學
【學位級別】：碩士
【學位授予年份】：2016
【分類號】：TN912.3;TP391.3
【參考文獻】
相關期刊論文 前4條
1 王恩成;蘇騰芳;袁開國;伍淳華;王玉慶;;哼唱檢索中聯合音高與能量的音符切分算法[J];計算機工程;2012年09期
2 毛大偉;曹華;木拉提.哈米提;童勤業(yè);;基于美爾倒譜系數和復雜性的說話人識別[J];生物醫(yī)學工程學雜志;2006年04期
3 劉建;鄭方;吳文虎;;基于幅度差平方和函數的基音周期提取算法[J];清華大學學報(自然科學版);2006年01期
4 徐險峰;基于內容的多媒體信息檢索技術[J];現代情報;2005年03期
，本文編號：2180532
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2180532.html
