New Artificial Voice Box Uses Synthesizer and Speech Recognition

A new type of artificial larynx is being developed by researchers at the the University of the Witwatersrand in Johannesburg, South Africa. The new voice box will do away with the robotic, harsh sounding speech produced by current devices, such as that used by physicist Stephen Hawking. The system senses contact between the palate and tongue to figure out which word is being mouthed, with the aid of speech recognition software, and generates audio using a speech synthesizer.

"All of the currently available devices produce such bad sound--it either sounds robotic or has a gruff speaking voice," says Megan Russell, Witwatersrand PhD candidate. "We felt the tech was there for an artificial synthesized voice solution." Around 10,000 Americans every year are diagnosed with laryngeal cancer, according to the National Cancer Institute; most patients with advanced cases will have their larynx removed. The new system in development is notable in that it does not require surgical implantation.

The system uses a sensing device that looks like an orthodontic plate, normally used for speech therapy. The device, called a palatometer, is manufactured by CompleteSpeech, and tracks mouth movements by sensing contact between the tongue and palate with 118 embedded touch sensors. Special software for the artificial larynx system was written by Russell and her colleagues at the University. The system then attempts to translate the mouth movements into words then reproduces them on a small portable sound synthesizer.

Like any speech recognition software, it must be trained, and Russell has so far trained the system to recognize 50 common English words. Word and morpheme information can be represented on a binary space-time graph and entered into a database. Whenever the user speaks, contact patterns are compared against the database to recognize the corresponding word.

When various recognition algorithms are combined, the system can identify the correct word 94.14 percent of the time, not including words that the system classifies as "unknown" and chooses to skip. Since choosing the wrong word "could lead to some very difficult social situations," Russell says, it is preferable to have the system omit unclear words and remain silent.