基于深度学习的小尺度单元拼接语音合成方法研究

发布时间：2018-01-02 19:31

本文关键词：基于深度学习的小尺度单元拼接语音合成方法研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：语音合成技术旨在实现文本等输入信息到语音波形的转换。统计参数语音合成方法以及单元挑选与波形拼接方法是现阶段实现语音合成的两类主流方法。前者具有系统自动构建、合成语音平滑流畅等优点。但是受制于参数合成器等因素,其合成语音的自然度仍不够理想。在统计声学模型指导下使用帧级长度的小尺度单元进行单元挑选和波形拼接,是一种改进统计参数方法合成语音自然度的有效途径。传统小尺度单元挑选方法使用隐马尔科夫模型(hidden Markov model,HMM)进行声学建模和代价函数计算。而近年来,以深度神经网络为代表的深度学习方法已经在统计参数语音合成的声学建模中体现出了相对HMM的性能优势。因此,本文围绕基于深度学习的小尺度单元拼接语音合成方法开展研究工作。一方面,研究了用于指导小尺度单元挑选的神经网络声学建模方法,通过使用深度神经网络与递归神经网络等模型结构,提升了传统HMM模型的建模精度与合成语音质量;另一方面,提出了一种结合单元挑选和参数生成的语音合成方法,利用帧拼接方法实现了激励特征波形的生成,改善了传统统计参数合成方法中对于相位等激励信息建模能力的不足,提高了合成语音自然度。本文的研究工作具体如下:首先,本文提出了基于深度神经网络的帧拼接语音合成方法。该方法使用深度神经网络构建声学模型用于帧挑选中的目标代价与连接代价的计算,相对传统HMM模型提高了模型的预测精度与合成语音的主观质量。其次,本文研究了基于递归神经网络的小尺度单元挑选与波形拼接合成方法。该方法一方面采用结合长短时记忆单元的递归神经网络进行声学建模以改善深度神经网络的时序建模能力,另一方面引入多帧挑选策略以减少拼接点,取得了比基于深度神经网络的帧拼接方法更好的合成语音自然度。最后,本文设计实现了结合单元挑选激励生成的参数合成方法。该方法对提取的激励特征波形进行参数表征和声学建模,在合成阶段使用帧拼接方法生成激励特征波形的高频成分,同时使用参数生成方法预测滤波器特征,最终通过滤波合成语音波形。实验结果表明了该方法在改善统计参数方法合成语音自然度上的有效性。
[Abstract]:Speech synthesis technology aims to realize the conversion of input information such as text to speech waveform. Statistical parameter speech synthesis method, unit selection and waveform splicing method are two main methods to realize speech synthesis at present. System built automatically. Synthesis speech smooth and smooth and other advantages, but limited by the parameter synthesizer and other factors. The naturalness of the synthesized speech is still not ideal. Under the guidance of the statistical acoustic model, small scale units of frame length are used for unit selection and waveform stitching. It is an effective way to improve the statistical parameter method for speech naturality synthesis. Traditional small scale unit selection method uses hidden Markov model. HMMs are used for acoustic modeling and cost function calculation. The depth learning method represented by depth neural network has shown the performance advantage of HMM in the acoustic modeling of statistical parameter speech synthesis. On the one hand, the neural network acoustic modeling method used to guide the selection of small scale units is studied. By using depth neural network and recurrent neural network, the modeling accuracy and synthetic speech quality of traditional HMM model are improved. On the other hand, a speech synthesis method combining unit selection and parameter generation is proposed, and the excitation feature waveform is generated by frame splicing. Improve the traditional statistical parameter synthesis method for phase and other excitation information modeling ability, improve the synthesis speech naturalness. The research work in this paper is as follows: first. In this paper, a method of frame mosaic speech synthesis based on depth neural network is proposed, which uses depth neural network to construct acoustic model to calculate target cost and connection cost in frame selection. Compared with the traditional HMM model, the prediction accuracy of the model and the subjective quality of synthesized speech are improved. Secondly. In this paper, the method of small scale unit selection and waveform splicing synthesis based on recurrent neural network is studied. On the one hand, the acoustic modeling based on recurrent neural network combined with long and short memory unit is used to improve the depth neural network. The temporal modeling capability of. On the other hand, the multi-frame selection strategy is introduced to reduce the stitching points, and better synthetic speech naturalness is obtained than the frame stitching method based on depth neural network. Finally. In this paper, we design and implement the method of parameter synthesis based on unit selection excitation, which is used for parameter representation and acoustic modeling of the extracted excitation waveform. In the synthesis phase, frame splicing method is used to generate the high-frequency components of the excitation characteristic waveform, and the parameter generation method is used to predict the filter features. Finally, the speech waveform is synthesized by filtering. The experimental results show that the proposed method is effective in improving the speech naturalness of the statistical parameter method.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN912.3

【相似文献】