RI: Small: A New Voice Source Model: From Glottal Areas to Better Speech Synthesis

这项研究的目标是开发和评估一种新的VoiceSource模型，该模型基于对30名成年说话者声带的生理观察。现有来源模型的缺点可以部分归因于它们的开发方式：基于少数发言者的有限数据，没有直接的生理观察，也没有感知验证。更大的数据集不仅有助于开发可以解释说话人内部和跨说话人的一系列语音质量的源模型，而且还有助于理解如何以及哪些模型参数(S)是特定于说话人和/或性别的。模型开发将从最早的阶段考虑模型参数的感知效果。更好的源模型也可能提高语音处理算法的性能，如文本到语音合成(TTS)。通常在这种算法的开发中，重点一直放在与语音频谱包络相关的声学特征上。另一方面，声源的声学受到的关注较少。所提出的工作包括：1)记录声带振动的高速图像和15名男性和15名女性说话人的同时录音；2)从图像中提取声门区域函数来参数化新的声源模型；3)进行感知实验以发现哪些模型参数在感知上是显著的；以及4)在TTS中使用新的声源模型。该项目的跨学科团队(在建模、合成、识别、语音学和心理语言学方面拥有专业知识)是唯一有资格进行这项变革性研究的人。

The goal of the proposed research is to develop and evaluate a new voicesource model based on physiological observations of the vocal folds of 30 adult speakers. Shortcomings of existing source models can be in part attributed to the way in which they were developed: based on limited data from a few speakers, without direct physiological observations, and without perceptual validation. A larger dataset would help in not only developing a source model that could account for a range of voice qualities within and across speakers, but also result in an understanding of how and which model parameter(s) are speaker and/or gender specific. Model development will consider the perceptual effects of the model's parameters from the earliest stages.A better source model might also improve the performance of speech processing algorithms such as text-to-speech synthesis (TTS). Typically in the development of such algorithms, the emphasis has been on acoustic features related to the speech spectral envelope. The acoustics of the voice source, on the other hand, have received less attention. The proposed work involves: 1) recording high-speed images of vocal foldvibrations with simultaneous audio recordings from 15 male and 15 female speakers, 2) extracting glottal area functions from the images to parameterize a new voice source model, 3) performing perception experiments to uncover which model parameters are perceptually salient, and 4) using the new voice source model in TTS. The project's interdisciplinary team (with expertise in modeling, synthesis, recognition, phonetics, and psycholinguistics) is uniquely qualified to conduct this transformative research.