A Tongue-Tracking Artificial Larynx

Researchers in South Africa are working on a new kind of artificial larynx that won't have the raspy voice of existing devices. The system tracks contact between the tongue and palate to determine which word is being mouthed, and uses a speech synthesizer to generate sounds.

According to the National Cancer Institute, some 10,000 Americans are diagnosed with laryngeal cancer each year, and most patients with advanced cancer must have their voice box removed.

"All of the currently available devices produce such bad sound--it either sounds robotic or has a gruff speaking voice," says Megan Russell, a PhD candidate at the University of the Witwatersrand in Johannesburg, South Africa. "We felt the tech was there for an artificial synthesized voice solution."

The system uses a palatometer: a device that looks much like an orthodontic plate and is normally used for speech therapy. The device, made by CompleteSpeech of Orem, UT, tracks contact between the tongue and palate using 118 embedded touch sensors. The software for the artificial larynx was written by Russell and colleagues at the University of the Witwatersrand. Their work is being presented at the International Conference on Biomedical and Pharmaceutical Engineering this week in Singapore.

To use the device, a person puts the palatometer in her mouth and mouths words normally. The system tries to translate those mouth movements into words before reproducing them on a small sound synthesizer, perhaps tucked into a shirt pocket.

So far, Russell has trained the system to recognize 50 common English words by saying each word multiple times with the palatometer in her mouth. The information can be represented on a binary space-time graph and put into a database. Each time the user speaks, the contact patterns are compared against the database to identify the correct word.

Russell's team has tested the word-identification system using a variety of techniques. One approach involves aligning and averaging the data produced while training the device for a few instances of a word to create a template for comparison. Another compares features such as the area of the data plots on the graph, and the center of mass on the X and Y axes. A voting system compares the results of selected methods to see whether there is agreement. The researchers have also tested a predictive-analysis system, which considers the last word mouthed to help determine the next.

Russell says that when the voting and predictive elements are combined, the system identifies the correct word 94.14 percent of the time, although this doesn't include words that the system classifies as "unknown" and chooses to skip. Russell says that happens about 18 percent of the time. But choosing the wrong word "could lead to some very difficult social situations," Russell says, so it's best for the system to reject unclear words and remain silent.