You may think that Stephen Hawking's synthesized voice sounds a little ... unusual, but imagine how much weirder it would be to witness a child using that same adult voice to communicate. For many children who are unable to speak, however, they have no choice but to use assistive devices that utilize just such a voice. Now, help may be on the way. Norwegian researchers have developed a new method of creating synthetic speech, that actually sounds like it is being spoken by a child. Such technology could also allow computers to better recognize words spoken to them by young users.
One of the systems is the result of a collaboration between software company Lingit, and Media LT, a company that develops devices for assisted living.
Initially, the researchers created a master voice, which was made by combining recordings of multiple adult speakers reciting several thousand phrases - enough to create a workable library of words and sounds. Then, they recorded a single child reciting a smaller number of phrases, which were selected to include the sounds that are most essential to the Norwegian language.
When a computer compared the master voice to the child's voice, using the phrases as a point of reference, it was able to alter the master voice to make it sound like that of a child. "The result sounds rather like a child with unusual elocution skills, but it's still much better than the voice of an adult," said Lingit's Dr. Torbjørn Nordgård.
Over at the Norwegian University of Science and Technology, meanwhile, synthetic children's speech is being used to teach computer voice recognition systems to better understand the voices of children.
Presently, most voice recognition systems are tailored toward adult speech, and often have difficulty recognizing words spoken by younger users. In order for these systems to get the hang of children's voices, they would need to be "trained" on recordings of children speaking - recordings that aren't nearly as plentiful as those of adult speech.
In order to remedy this situation, the researchers created a synthetic child's voice of their own. In their case, they analyzed how children's shorter vocal tracts affect the frequency distribution of their speech energy. They then altered the energy distribution of an adult speech program, to achieve a child-like sound. "We could apply our conversion technique to a large database of adult speech and generate a functional database of artificial childlike voices," explained Prof. Torbjørn Svendsen. "We then used this to train a separate speech recognition program for children."
When tested, a prototype version of the program had an error rate 50 to 70 percent lower than traditional "adult-oriented" programs.
Source: The Research Council of Norway