Voice Montage

本研究的目的是综合考察语音的基本属性，特别是在1998-1999年“探索性研究资助”项目的成果基础上，寻找实现“语音蒙太奇”的途径。结果如下所示。2000年，人们设计并实施了几个实验，以假设一个人是否能记住别人的声音质量。在这些实验中，使用过去研究中建立的与语音质量相关的日常表达来评估他人声音的听觉印象。因此，我们确定了实现声音蒙太奇的大前提：“人类有能力在一定时间内相当精确地保持对听到的声音质量的记忆，即使是对未知的声音。”2001年，声学参数与语音质量相关表达之间的关系被研究，这对于从听觉印象合成语音具有重要意义。首先，采用图形化建模方法研究了更多的声学参数。其次，利用非线性决策树对听觉实验中声学参数与评价之间的关系进行建模。因此，结果似乎比线性多元回归的结果更合理。例如，非线性分析表明，在描述“高音声音”时，语速是仅次于基频的第二重要因素。此外，我们已经开始构建与语音传递的准语言/语言外信息相关的表达集。这些表达包括说话风格和反映听者价值观的表达，以及之前研究中提取的表达。通过以下方法收集语音相关表达词。几名年龄在20岁至60岁之间的工作人员被要求仔细扫描一本日语词典的所有条目，以找到与语音相关的单词。通过这项工作，收集了75000个单词中的3076个。然后，通过考虑一致性和明显的同义性，将这些单词过滤并合并为872个单词。最后，将这些词分为四组，即与“语音质量”、“说话风格”、“说话风格导向的语音质量”和反映听者价值观的词。这些表情被认为包含了除语音质量之外的个性信息。少

The goal of this research is to find a way to realize "voice montage" by comprehensively investigating fundamental properties of voice, especially based on the past achievement of the "grant-in-aid for exploratory research" project in 1998-1999. The results are described below.In 2000, several experiments were designed and executed to hypothesize whether one can remember the others' voice quality. In these experiments, the auditory impression of the others' voice was evaluated with the voice-quality-related everyday expressions which were established in the past research. As a result, we confirmed the following major premise for realizing voice montage : "Man has the ability to fairly precisely maintain the memory of heard voice's quality for a certain period of time, even for unknown voice."In 2001, the relationship between acoustic parameters and voice-quality-related expressions was examined, which would be important for synthesizing voice from the auditory impression. First, depend … More ence of the acoustic parameters were investigated using graphical modeling. Next, the relationship between the acoustic parameters and the evaluation in auditory experiments was modeled by the non-linear decision tree. Consequently, the result seems to be more reasonable than that of linear multiple regression. For example, the non-linear analysis indicates that speaking rate is the second most important factor next to fundamental frequency for describing "high-pitched voice."In addition, we have started the construction of the set of expressions related to para-linguistic/extra-linguistic information which is conveyed by voice. The expressions include speaking style and the ones which reflect hearers' sense of value, as well as the ones extracted in the previous research. The voice-related expression words were collected with the following method. Several workers aged from 20s to 60s were instructed to find voice-related words by carefully scanning all entries of a Japanese dictionary. 3,076 words out of 75,000 were collected through the work. Those words were then filtered and merged into 872 words by considering the agreement and apparent synonymity. Finally, the words were divided into four groups, i.e. the ones related to "voice quality," "speaking style," "speaking-style-oriented voice quality," and the ones which reflect hearers' sense of value. The expressions are thought to contain individuality information other than voice quality. Less