Instrumental Quality Estimation for Synthesized Speech Signals

在最初的提案中计划了三年的项目[穆勒，豪特，2009]，到目前为止已经批准了两年，目的是对文本到语音(TTS)合成系统的听觉质量进行工具测量。因此，昂贵的听觉测量(听力测试)不应完全被取代；但是，在系统开发或预选潜在的候选系统期间，它们可以被广泛避免。目标是一种仅根据TTS合成信号来估计质量的度量，而没有任何参考信号。它是以属性为基础的，即它像听众一样，从不同的单一方面汇编总体印象。因此，感知质量空间被一定数量的维度所跨越。对于尺寸，需要找到仪器测量，最后，从这些测量中构建一个完整的质量测量。然而，与直接的整体质量评估不同的是，详细的诊断信息是可用的，对于系统开发人员来说是有价值的。在最初的提案中，提出了以下问题：在不同的应用中，哪些方面与合成语音的感知和质量判断相关？这些方面中的哪些反映在语音信号中？如何才能对单一质量方面进行工具性评估？如何仅从合成的语音信号来估计总体质量？如何在不同应用程序的质量配置文件中表示估计值？对于哪些合成方法和应用来说，这样的概况是成立的？经过头两年，已经找到了部分答案。为了继续调查，增加了以下问题：测量的质量如何取决于TTS扬声器的声音？如何将超细分市场的质量因素考虑在内？一方面，在时间变化的意义上，相关特征如何依赖于时间；另一方面，在合理的估计可靠性的意义上，相关特征如何依赖于时间？哪些特征对短TTS信号的质量分析有用，哪些特征用于较长序列的质量分析？虽然从单一方面汇编整体质量以及有效性问题在最初的提案中主要计划在第三年进行[Möller，Heant，2009]，但由于目前的调查，对纳入声音质量和时间相关性的扩展是新的结果。同样，超节段质量测量的核心重要性是一个新的观点，现在要考虑到。

The project planned for three years in the initial proposal [Möller, Heute, 2009], granted up to now for two years aims at an instrumental measure for the auditory quality of text-to-speech (TTS) synthesis systems. Thereby, expensive auditive measurements (listening tests) shall not be replaced completely; they can, however, be avoided widely during a system development or pre-selection of potential system candidates. The goal is a measure which estimates the quality from the TTS-synthesis signal only, without any reference signal. It is based on attributes i.e., it compiles the total impression, like a listener, from different single aspects. So, the perceptual quality space is spanned by a certain number of dimensions. For the dimensions, instrumental measurements are to be found, from which, finally, an integral quality measure is to be constructed. Beyond, however, and different from a direct integral quality estimation, detailed diagnostic information is available, valuable for the system developer.In the initial proposal [Möller, Heute, 2009] the following questions were raised: Which aspects are relevant for the perception and quality judgment of synthetic speech at various applications? Which of these aspects are reflected in the speech signal? How can the single quality aspects be estimated instrumentally? How can the total quality be estimated from the synthesized speech signal alone? How can the estimated values be represented in a quality profile for different applications? For which synthesis methods and applications does such a profile hold?After the first two years, partial answers have been found. For the continued investigations, the following questions are added: How does the measured quality depend on the TTS speakers voice? How can supra-segmental quality aspects be taken into account? How do relevant features depend on time in the sense of time-variance on one hand, in the sense of a reasonable estimation reliability on the other hand? Which features are useful for the quality analysis of short TTS signals, which ones for longer sequences? While the compilation of the integral quality from single aspects as well as the validity question were already planned mainly for the third year in the initial proposal [Möller, Heute, 2009], the extension towards an inclusion of voice quality and time dependence resulted newly from the present investigations. As well, the central importance of a supra-segmental quality measurement is a novel point of view, now to be taken into account.