韵律边界标注对于语料库建设和语音合成有着至关重要的作用，而自动韵律标注可以克服人工标注中耗时、不一致的缺点。仿照人工标注流程，该文运用循环神经网络分别对文本和音频两个通道训练子模型，对子模型的输出采用模型融合的方法，从而获得最优标注。以词为单位提取了静音时长，与传统以帧为单位的声学特征相比更具有明确的物理意义，与韵律边界的联系更加紧密。实验结果表明：相比传统声学特征，该文所采用的静音时长特征使自动韵律标注的性能有所提高；相比直接特征层面的方法，决策融合方法更好地结合了声学和文本的特征，进一步提高了标注的性能。
Abstract
Automatic prosodic boundary labeling is important in the construction of a speech corpus for speech synthesis. Automatic labeling of prosodic boundaries gives more consistent results than manual labeling of prosodic boundaries which is time consuming and inconsistent. Manual labeling method is modelled here using a recurrent neural network to train two sub-models which use lexical features and acoustic features to label the prosodic boundaries. Model fusion is then used to combine the outputs of the two sub-models to obtain the optimal labeling results. The silence durations for each word give clearer physical meanings and better correlations with the prosodic boundaries than the acoustic features used in traditional methods extracted frame-by-frame. Tests show that the silence durations extracted using the current acoustic features and the model fusion method improve the prosodic boundary labeling compared with previous feature fusion methods.
关键词
韵律边界标注 /
决策融合 /
静音时长 /
语料库构建 /
语音合成
Key words
prosodic boundary labeling /
ensemble strategy /
silence duration /
corpus construction /
speech synthesis
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] CHU M, QIAN Y. Locating boundaries for prosodic constituents in unrestricted Mandarin texts[J]. Computational Linguistics and Chinese Language Processing, 2001, 6(1):61-82.[2] WANG M Q, HIRSCHBERG J. Automatic classification of intonational phrase boundaries[J]. Computer Speech & Language, 1992, 6(2):175-196.[3] LEVOW G A. Automatic prosodic labeling with conditional random fields and rich acoustic features[C]//International Joint Conference on Natural Language Processing (IJCNLP). Hyderabad, India:2008:217-224.[4] ROSENBERG A, FERNANDEZ R, RAMABHADRAN B. Modeling phrasing and prominence using deep recurrent learning[C]//Proceedings of the Annual Conference of the International Speech Communication Association (INTERSPEECH). Dresden, Germany, 2015:136-141.[5] BUSSER B, DAELEMANS W, BOSCH A. Predicting phrase breaks with memory-based learning[C]//4th ISCA Tutorial and Research Workshop (ITRW) on Speech Synthesis. Edinburgh, UK:University of Edinburgh, 2001:29-34.[6] WIGHTMAN C W, OSTENDORF M. Automatic labeling of prosodic patterns[J]. IEEE Transactions on Speech and Audio Processing, 1994, 2(4):469-481.[7] HASEGAWA-JOHNSON M, CHEN K, COLE J, et al. Simultaneous recognition of words and prosody in the boston university radio speech corpus[J]. Speech Communication, 2005, 46(3):418-439.[8] CHEN Q, LING Z H, YANG C Y, et al. Automatic phrase boundary labeling of speech synthesis database using context-dependent HMMs and N-Gram prior distributions[C]//Proceedings of the Annual Conference of the International Speech Communication Association (INTERSPEECH). Dresden, Germany, 2015:227-234.[9] DING C, XIE L, YAN J, et al. Automatic prosody prediction for Chinese speech synthesis using BLSTM-RNN and embedding features[C]//Automatic Speech Recognition and Understanding (ASRU). Scottsdale, USA, 2015:98-102.[10] LIN C K, LEE L S. Improved spontaneous Mandarin speech recognition by disfluency interruption point (IP) detection using prosodic features[C]//Ninth European Conference on Speech Communication and Technology. Lisbon, Portuguese, 2005:78-85.[11] TIELEMAN T, HINTON G. Lecture 6.6-Rmsprop:Divide the gradient by a running average of its recent magnitude[Z/OL].[2017-01-01]. https://www.coursera.org/learn/neural-networks.[12] HINTON G E, SRIVASTAVA N, KRIZHEVSKY A, et al. Improving neural networks by preventing co-adaptation of feature detectors[J]. Computer Science, 2012, 3(4):212-223.
PDF(1161 KB)
