Models of speech synthesis.

Abstract

The term "speech synthesis" has been used for diverse technical approaches. In this paper, some of the approaches used to generate synthetic speech in a text-tospeech system are reviewed, and some of the basic motivations for choosing one method over another are discussed. It is important to keep in mind, however, that speech synthesis models are needed not just for speech generation but to help us understand how speech is created, or even how articulation can explain language structure. General issues such as the synthesis of different voices, accents, and multiple languages are discussed as special challenges facing the speech synthesis community.The term "speech synthesis" has been used for diverse technical approaches. Unfortunately, any speech output from computers has been claimed to be speech synthesis, perhaps with the exception of playback of recorded speech.* Some of the approaches used to generate true synthetic speech as well as high-quality waveform concatenation methods are presented below.Knowledge About Natural Speech Synthesis development can be grouped into three main categories: acoustic models, articulatory models, and models based on the coding of natural speech. The last group includes both predictive coding and concatenative synthesis using speech waveforms. Acoustic and articulatory models have had a long history of development, while natural speech models represent a somewhat newer field. The first commercial systems were based on the acoustic terminal analog synthesizer. However, at that time, the voice quality was not good enough for general use, and approaches based on coding attracted increased interest. Articulatory models have been under continuous development, but so far this field has not been exposed to commercial applications due to incomplete models and high processing costs.We can position the different synthesis methods along a "knowledge about speech" scale. Obviously, articulatory synthesis needs considerable understanding of the speech act itself, while models based on coding use such knowledge only to a limited extent. All synthesis methods have to model something that is partly unknown. Unfortunately, artificial obstacles due to simplifications or lack of coverage will also be introduced. A trend in current speech technology, both in speech understanding and speech production, is to avoid explicit formulation of knowledge and to use automatic methods to aid the development of the system. Since such analysis methods lack the human ability to generalize, the generalization has to be present in the data itself. Thus, these methods need large amounts of speech data. Models working close to the waveform are now typically making use of increased unit sizes while still modeling prosody by rule. In the middle of the scale, "formant synthesis" is moving toward the articulatory models by looking for "higher-level parameters" or to larger prestored units. Articulatory synthesis, hampered by lack of data, still has some way to go but is yielding improved quality, due mostly...