Knowledge from Speech Production Used in Speech Technology: Articulatory Synthesis*

There appears to be a continuing trend toward incorporating knowledge of speech production into s~eech technology-text-to-speech synthesis (e.g., BIckley, Stevens, & Williams, 1994; Parthasarthy & Coker, 1992), low bit rate coding (see Schroeter & Sondhi, 1992), and automatic speech recognition (e.g., Rose, Schroeter, & Sondhi, 1994; Shirai & Kobayashi, 1986). For automatic speech recognition, using knowledge of the coordination of the vocal tract articulators and the resulting acoustics can reduce apparent token-to-token variability so that general pattern recognition alg.orithms have less work to do. Using artIculatory representations in speech coding has the potential of greatly reducing bit rate because the articulators move relatively slowly and may be described by a few parameters by using an underlying dynamical model or by using simple curve fitting. Finally, text-to-speech synthesis can be improved using articulator control parameters, because the laws ofphysics can be used to produce the correct bundle of acoustic features with a comparatively limited parameterization-the acoustic output is constrained by the laws of physics. All these applications that depend on articulatory representation of speech production, can be grounded in what is called an articulatory synthesizer. An articulatory synthesizer is a device that produces speech output from a set of articulatory parameters (an articulatory representation). These devices are usually implemented in software on a digital computer.