数字音乐语音智能评判方法研究与实现

发布时间：2018-05-04 00:07

本文选题：数字音乐语音 + 哼唱识别　；参考：《福州大学》2014年硕士论文

【摘要】：随着音乐娱乐的普及和发展,数字音乐语音智能处理研究受到广泛关注。音乐演唱的音准音高智能评分,已经在不少音乐娱乐软件产品中推广应用。但针对音乐语音的哼唱识别、演唱的情感识别等,尚处于起步阶段,还需要进一步扩展计算机智能评判的方法。相较普通说话语音,数字音乐语音具有单字发音时长变化范围广、发音受歌曲节奏影响大、语音处理的实时性要求高等特点。传统语音处理方法在数字音乐的应用效果并不理想,如音乐语音分割无法将单字准确切分开,传统语音处理选取特征的方法无法有效适用于音乐语音识别。针对单字发音时长变化范围广的特点,本文提出了基于信息熵的哼唱识别算法和基于粒子群优化的数字音乐语音分割算法。信息熵分割算法对于哼唱时发音无停顿的特点十分敏感,利用该算法能够准确检测出在KTV唱歌的过程中有人利用哼唱获取高分的情况；传统端点检测技术难以适应数字音乐语音节奏变化快、发音间隔不规律、字与字之间的连续无停顿或停顿间隔不明显等特点,因此本文提出了利用粒子群优化对未完全分割开的多个字进行细分的方法,该方法能够智能地判断一个长语音段所包含的字数并准确地找到每一个字发音的起始位置和结束位置,在有噪音干扰的条件下同样适用。实验结果表明,上述两种算法能够有效准确处理数字音乐语音,提高音乐语音的分割准确率。目前对于语音情感识别的研究没有找到某种特征能够非常典型的、单独表示和区分不同情感,因此需要从现有已知的特征中选择合适的特征用于情感识别。本文提出了一种基于DTW系数与SVM投票机制的数字音乐语音情感识别算法,该算法考虑音乐语音的音质特征和韵律特征,利用DBC系数从备选特征中提取出类间相似度小、类内相似度大的特征,然后建立不同的二分SVM分类器,对每一个样本进行分类决策,最后统计每个分类器的结果,得票数最多的情感即为识别结果。通过与Fisher准则提取特征的SVM分层语音情感识别算法的对比试验表明,由于消除了不同情感样本数不同对Fisher投影区域的影响及分层算法累积误差的影响,本文提出的基于DBC系数与SVM投票机制的数字音乐语音情感识别算法有效地提高了识别准确率。最后,本文将上述算法实现并封装成模块加入KTV评分系统。本文考虑数字音乐语音与正常说话的不同,提出的音乐语音哼唱检测、分割算法和情感识别算法弥补传统算法的不足,具有较高的研究价值和广阔的发展前景。
[Abstract]:With the popularity and development of music entertainment, the research of digital music speech intelligent processing has received extensive attention. Music singing pitch intelligence score, has been used in many music entertainment software products. However, the humming recognition of musical speech and the emotion recognition of singing are still in the initial stage, and the methods of computer intelligent evaluation need to be further expanded. Compared with ordinary speech speech, digital music speech has the characteristics of wide range of changes in the duration of single word pronunciation, great influence of song rhythm on pronunciation, and high real-time requirement of speech processing. The application effect of traditional speech processing method in digital music is not ideal, such as music speech segmentation can not accurately separate words, traditional speech processing method of feature selection can not be effectively applied to music speech recognition. Aiming at the wide range of pronunciation time, this paper proposes a humming recognition algorithm based on information entropy and a digital music speech segmentation algorithm based on particle swarm optimization (PSO). The information entropy segmentation algorithm is very sensitive to the feature of no pause in the pronunciation of humming. Using this algorithm, it can accurately detect the situation that someone gets high score by humming in the process of singing by KTV. The traditional endpoint detection technique is difficult to adapt to the characteristics of fast change of digital music voice rhythm, irregular pronunciation interval, continuous non-pause between words or not obvious pause interval, etc. In this paper, a particle swarm optimization (PSO) method is proposed to subdivide several words with incomplete segmentation. This method can intelligently judge the number of words contained in a long speech segment and accurately find the starting and ending position of each word, which is also applicable in the case of noise interference. The experimental results show that the two algorithms can deal with digital music speech accurately and improve the segmentation accuracy of music speech. At present, the research on speech emotion recognition has not found that a certain feature can be very typical, which can express and distinguish different emotions separately, so it is necessary to select suitable features from existing known features for emotion recognition. In this paper, an emotion recognition algorithm for digital music speech based on DTW coefficient and SVM voting mechanism is proposed. The algorithm takes into account the sound quality and prosodic features of music speech, and uses DBC coefficient to extract small similarity between classes from the alternative features. Then different binary SVM classifiers are established to make classification decisions for each sample. Finally, the results of each classifier are counted, and the emotion with the most votes is the recognition result. By comparing with the SVM hierarchical speech emotion recognition algorithm based on Fisher criterion, it is shown that the influence of different number of emotional samples on Fisher projection region and the cumulative error of layered algorithm are eliminated. An emotion recognition algorithm for digital music speech based on DBC coefficient and SVM voting mechanism is proposed in this paper. Finally, the above algorithm is implemented and encapsulated into a KTV scoring system. Considering the difference between digital music speech and normal speech, this paper puts forward the music speech humming detection, segmentation algorithm and emotion recognition algorithm to compensate for the shortcomings of the traditional algorithm, which has high research value and broad development prospects.
【学位授予单位】：福州大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.3

【相似文献】