小波域多重分形和能量谱参数的病理嗓音识别研究

发布时间：2018-06-23 00:13

本文选题：病理嗓音识别 + 小波leaders多重分形　；参考：《苏州大学》2016年硕士论文

【摘要】：随着社会的发展,人们之间的交往日益频繁,语言是人与人交流的重要桥梁,但是嗓音疾病导致嗓音嘶哑、无力等,严重影响了人们的生活质量和社会交流。采用声学分析技术对病理嗓音信号进行研究和分析,可以实现对嗓音质量的客观评估,对喉部疾病的诊断和治疗具有重要意义。小波变换能在不同尺度对信号进行分析,多重分形谱能够精细刻画信号局部尺度行为,同时小波变换是一种时频分析方法,能有效反映嗓音在时间尺度平面能量分布,因此本文主要研究了嗓音小波域多重分形谱参数和能量谱参数识别病理嗓音。针对单一分形维数在描述嗓音非线性方面的不足,引入分形和小波结合的小波leaders多重分形谱,并采用Chhabza算法求解小波leaders多重分形谱,通过非参数bootstrap统计估计参数置信区间。小波leaders多重分形谱能够表征嗓音信号的局部奇异性统计分布情况,其参数在正常嗓音和病理嗓音间具有较大差异,病理嗓音谱宽较小,正常嗓音谱宽较大,正常嗓音多重分形特征更加明显,小波leaders多重分形谱参数对病理嗓音的平均识别率为90.66%。传统声学参数无法很好的表征信号非平稳特性,本文对嗓音在小波域进行时频分析,提出一种小波域能量谱参数GCWT,对病理嗓音进行识别。对沿尺度轴方向的能量谱进行多维高斯混合建模,并且将模型参数作为嗓音特征参数GCWT进行病理嗓音识别,相对于传统参数,参数GCWT对病理嗓音具有比较好的识别效果,平均识别率为92.99%。由于参数GCWT维数较高,采用主成分分析和局部线性嵌入对特征参数GCWT进行降维处理,并且提出了动态加权局部线性嵌入DWLLE降维方法,弱化稀疏采样对降维的影响,有效保留了参数高维空间的几何特性,降低了参数冗余,降维后的参数对病理嗓音的识别率最高达97.45%。
[Abstract]:With the development of society, communication between people is becoming more and more frequent. Language is an important bridge between people, but voice diseases lead to hoarseness and powerlessness, which seriously affect people's quality of life and social communication. Using acoustic analysis technology to study and analyze pathological voice signals can realize the objective evaluation of voice quality and has important significance for the diagnosis and treatment of laryngeal diseases. Wavelet transform can analyze the signal at different scales, multifractal spectrum can describe the local scale behavior of the signal in detail, and wavelet transform is a time-frequency analysis method, which can effectively reflect the energy distribution of voice in time scale plane. In this paper, the multifractal and energy spectrum parameters of voice wavelet domain are studied to identify the pathological voice. Aiming at the shortage of single fractal dimension in describing voice nonlinearity, wavelet leaders multifractal spectrum combined with fractal and wavelet is introduced, and wavelet leaders multifractal spectrum is solved by Chhabza algorithm. Parameter confidence interval is estimated by nonparametric bootstrap statistics. Wavelet leaders multifractal spectrum can characterize the statistical distribution of local singularity of voice signal. The parameters of wavelet leaders multifractal spectrum are different between normal voice and pathological voice, the spectrum width of pathological voice is smaller, and the spectrum width of normal voice is larger. The multifractal feature of normal voice is more obvious, and the average recognition rate of wavelet leaders multifractal spectrum parameters for pathological voice is 90.66. The traditional acoustic parameters can not well characterize the non-stationary characteristics of the signal. In this paper, the time-frequency analysis of the voice in the wavelet domain is carried out, and a wavelet domain energy spectrum parameter GCWTis is proposed to identify the pathological voice. The multi-dimensional Gao Si hybrid modeling of energy spectrum along the scale axis is carried out, and the model parameters are used as the voice characteristic parameters to identify the pathological voice. Compared with the traditional parameters, the parameter Gao Si has a better recognition effect on the pathological voice. The average recognition rate was 92.99%. Due to the high dimension of parameter GCWT, the dimensionality reduction of characteristic parameter GCWT is performed by principal component analysis and local linear embedding, and a dynamic weighted locally linear embedding DWLLE method is proposed to reduce the dimension, which weakens the effect of sparse sampling on dimensionality reduction. The geometric characteristics of the high-dimensional space are effectively preserved and the redundancy of the parameters is reduced. The recognition rate of the parameters after dimensionality reduction is 97.45%.
【学位授予单位】：苏州大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：R767.92

【相似文献】