Synthesizing Vowel Transitions with an Analog Vocal Tract

Michael Brady - mcbrady@indiana.edu
Indiana University

Popular version of paper 1aSCa4
Presented Monday morning, May 18, 2009
157th ASA Meeting, Portland, OR

Virtually all research involving speech synthesis today relies on the computer to generate sound. Digital synthesis involves either cutting and splicing recorded segments of speech together or produces sound based on simulating the mechanics of the vocal tract. While these approaches work well for many lines of inquiry, speech production is increasingly being approached as a problem of motor control using techniques from robotics. These techniques typically involve feedback loops and reflex-like adaption. Digital simulation arguably does not capture the precise motor-to-sound mapping or the rich aerodynamics of speech that is called for in feedback-based motor control. The primary advantage of a mechanical or analog vocal tract is that it "computes the physics for free" - where computers simply are not fast or accurate enough to model how each particle of air moves around in the mouth when talking. Modern advances in robotics technologies (such as cheap, powerful, and accurate motors, sophisticated microcontrollers, and various other components like plastics and resins) allow for an analysis-by-synthesis approach to speech research in terms of true acoustics and aerodynamics.

In this talk I introduce a mechanical vocal tract that uses an air-based sound source with artificial vocal cords, silicone tongue, lip rounding mechanism, and nasal cavity. The tract is designed to make fluid movements, and the production targets for the tract are conceptualized as movements rather than as idealized static motor configurations. I discuss this "transition-versus-static target" distinction with a focus on vowel sounds as I consider some examples and review some experimental results. Videos of the tract are provided to further demonstrate the importance of controlled movements in speech and why the design of a mechanical tract needs to concentrate more on movements than on static positions to be achieved. In conclusion, I discuss the potential of the analog tract as a powerful emerging tool in speech research. For instance, current work on the tract involves mounting air pressure sensors along the roof of the vocal cavity and pumping smoke through the cavity during vocalization (the tract has see-through walls) to visualize and better appreciate changes in air turbulence during speech. It is not possible to conduct these kinds of investigations using the vocal tracts of people.

For further information, please visit the web site: http://www.fluidbase.com/mech_tract