Toshiba eyes auto apps for voice-recognition middleware
By Yoshiko Hara
EE Times
January 23, 2002 (5:45 p.m. EST)
TOKYO — Toshiba Corp. has developed middleware that can recognize voices in nine different languages, and plans to begin marketing it worldwide this spring for use in mobile and automotive applications.
The voice recognition middleware has two parts: the recognition processing section is not language dependent; the section that depends on acoustic and language models of the language to be recognized is. By combining multiple language-dependent sections with the core recognition algorithm, the system can recognize many languages. Toshiba has prepared phonemic models for Japanese, American English, British English, French, Germany, Spanish, Italian, Dutch and Chinese. The phonemic models of each language include some accent variations.
"Embedded software is required to be compact because CPU power and memory are limited. That's the challenge," said Hiroshi Kanazawa, senior research scientist at the multimedia laboratory of Toshiba's corporate research and development center in Kawasaki, Japan.
Multilanguage voice recognition systems, such as IBM's ViaVoice, are already available, but Kanazawa said the compactness of Toshiba's system meets a basic requirement of embedded applications. A standard configuration for a car navigation application would require about 1 Mbyte of ROM and 3 Mbytes of RAM. By minimizing the size of the acoustic model, a minimum configuration of about 500 kbytes could be implemented in a cellular phone with only a slight sacrifice of recognition capability, Kanazawa said.
The small memory requirement limits the number for words the system can recognize. With 3 Mbytes of RAM, it can recognize about 4,000 place names, Toshiba said.
Toshiba said it will supply and port the middleware to a customer's requested environment.
Aside from its compactness, the middleware is also differentiated by its ability to withstand extraneous noise, Kanazawa said. "We have established acoustic models which work well in various environments, from quite to very noisy, where the signal-to-noise ratio is about zero dB. This type of wide noise robustness is to our advantage."
Toshiba engineers said their system recognized about 90 percent of words on average when tested in a moving car, while competitors' systems recognized about 70 percent.
In-car applications are the middleware's initial target. "We believe the car will be a killer application for voice recognition systems," Kanazawa said.
Toshiba has already licensed the middleware to manufacturers of car navigation systems in Japan. The company's e-solutions company will begin sales promotion in Europe and the United States this spring, and in China in the fourth quarter.
Toshiba intends to brush up the middleware's phonemic models for each language at R&D locations in Yokohama, Japan, in Europe, and in China.
Toshiba engineers plan to present the technology this March at the spring meeting of the Acoustical Society of Japan, to be held March 18-20 at Kanagawa University in Yokohama.
AI
