Tokyo--Toshiba Corporation today announced a new generation of speech-recognition middleware that supports nine languages: Japanese, American and British English, French, German, Spanish, Italian, Dutch and Chinese. Expressly designed for compactness and high-speed execution, the middleware does not require a large memory capacity or high processing capability, and is ideally suited for integration in such applications as car navigation systems and mobile phones.

Spoken language is widely recognized as the ideal interface between people and machines, with the potential to surpass keyboard and pen-based input in convenience and flexibility. Speech recognition technology has already been developed for PCs and car navigation systems, and will soon allow users to issue commands to mobile phones and home appliances.

Practical voice recognition technology must achieve the indispensable target of high performance not only in quiet environments but also in places as noisy as a moving car or crowded shopping street. To achieve this, Toshiba developed original noise cancellation and noise adaptation techniques. The company's research into speech recognition has resulted in an advanced algorithm, "noise immunity training," that achieves high levels of recognition in real-world environments. A Japanese speech recognition system employing this algorithm has already been applied to car navigation systems.

The new middleware draws on Toshiba's work in Japanese speech recognition to deliver acoustic models for multiple languages. Each acoustic model was developed from a voice database of many speakers of different ages, voice registers and accents within the country where the languages is spoken. So far, nine acoustic models are available, and more will be developed in the coming year. Up to this point, the company has promoted speech recognition research at its corporate Research & Development Center in Kawasaki, Japan and the Toshiba China laboratory in Beijing, China. In the coming year, Toshiba will also found another research center in Europe that will work on improved speech recognition technology in European languages. Toshiba's goal is to deliver a versatile system that allows equipment makers to easily integrate recognition of multiple languages into their products.

Product Features

The new voice recognition middleware enjoys clear technical advantages.

Recognition of nine languages Toshiba's multi-lingual speech recognition middleware recognizes nine languages. The company has also developed nine voice databases for designing acoustic models that support recognition of regional pronunciations. The Chinese version was developed through collaboration by the company's Research & Development Center in Japan and Toshiba China laboratory.

Excellent Noise Tolerance Toshiba's proprietary technology, the "noise immunity training" algorithm, was used in development of the speech recognition engine.

Few Demands Processor Power and Memory Size The engine is designed to work with limited resources in terms of processing power and memory size. Performance is excellent with RISC chips.

Separation of Language-Processing Unit from Recognition-Processing Unit The language-processing unit is separated from the recognition-processing unit. This allows efficient substitution of one language for another, and also equips the language-processing unit to interface with other applications, such as speech synthesis and machine translation systems. The design concept supports product developers in the efficient construction of total system.

Platform independent The engine works with diverse operating systems and processors. It supports the μITORN, WinCE, Linux and pSOS OS and the TX, ARM and SH processor architectures.

Functionality Other features include a large vocabulary, word spotting and and recognition of connected digits.