Automatic target cost and database design for unit-selection speech synthesis

我们建议替换典型的串联语音合成器的三个组成部分：文本选择算法（为数据库记录什么），目标成本函数（从数据库中选择哪些单元）和后退策略（当数据库不包含所需单元时该怎么办）。目前，这些组件都是利用人类直觉独立设计的。这是非常困难的，只能由专家来完成，这意味着每个组件都不可能相对于其他组件进行优化。我们建议将这三个组件建立在一个底层模型上。该模型将从数据中学习哪些语音单位在感知上是可互换的。这些信息将被目标成本函数/撤退策略使用，并在选择要记录的文本时使用。提出的技术将在Festival 2语音合成系统中实施，并使用正式的听力测试进行评估。我们把研究计划分为三个阶段。在第一阶段，我们将对当前的技术有更深的理解。在第二阶段，我们将研究针对现有语音学习目标成本/后退策略的技术，然后针对给定目标成本/后退策略学习文本选择算法。最后，在第三阶段，我们将设计一种联合学习两者的方法。

We propose to replace three components of a typical concatenativespeech synthesiser: the text selection algorithm (what to record forthe database), the target cost function (which units to select fromthe database) and the backoff strategy (what to do when the databasedoes not contain the desired unit).These components are currently designed independently using humanintuition. This is very hard, can only be done by experts, and meansthat each component is unlikely to be optimised with respect to theothers. We propose to base these three components on a singleunderlying model. The model will learn, from data, which speech unitsare perceptually interchangeable. This information will then be usedby the target cost function / backoff strategy, and when selecting thetext to be recorded. The proposed techniques will be implemented inthe Festival 2 speech synthesis system and evaluated using formallistening tests.We break down the research programme into three phases. In Phase 1, wewill gain a deeper understanding of current techniques. In Phase 2, wewill examine techniques for learning just the target cost/backoffstrategy, given an existing voice, then for learning thetext-selection algorithm for a given target cost/backoffstrategy. Finally, in Phase 3, we will devise a method for jointlylearning both together.