Effects of Pause Insertion on the Intelligibility of Low Quality Speech

The intelligibility of the Output of text-to-speech Systems today is generally poorer than that of natural human speech. One way of improving the quality of eynthetic speech is to insert speech pauses at selected positions in the utterances rather more frequently than the human reader would choose to do. Pause insertion has been reported to improve intelligibility in deaf-speech [1] äs well äs in speech synthesized from diphones [2]. In the studies cit^d here, the pauses were inserted at breaks in the syntactic structure of the spoken sentences, without explicit motivation for this particular pausing strategy. Although the choice is intuitively appealing, there may be other pausing strategies that are perceptually äs adequate but easier to use in text-to-speech applications. The present research, therefore, aims to establish an optimal pause insertion strategy for low-quality speech synthesis. There are at least two criteria that have to be considered when choosing a pausing strategy. Firstly, the pauses should convey äs much useful Information to the listener äs possible; secondly, the positions where the pauses are to be inserted should be detectable by a simple algorithm that can easily be incorporated in a text-to-speech System. Generally, the more useful the Information signalled by the speech pause, the harder it is to find its Position automatically. In our experiments we Bystematically examined the effects of four pausing strategies, which will be discussed in our next sections in ascending order of complexity. The first strategy is one without pauses. The second strategy has pauses at word boundaries at regulär intervals of six words. Here the listener gets Information about some word beginnings in the sentence. In the third strategy we have marked word boundaries before Content words so that Information is given to the listener about beginnings of relatively important words in a sentence. In the last strategy we marked boundaries so äs to reveal the prosodic structure of the sentence. The speech pauses in this version are located at the end of intonational or phonological phrases. We expect this last strategy to be the most helpful to improve the intelligibility of low quality speech.