NEC Develops Speech-to-Speech Translation Software for Low Power Consumption Multi-Core Processors Optimal for Small Devices such as Mobile PhonesTokyo, October 24, 2005 --- NEC Corporation today announced that it has succeeded in the development of Japanese-English/English-Japanese, automatic speech translation software for single-chip multi-core processors for small devices such as mobile phones, capable of operation at high speeds with low power consumption. NEC verified the high-speed automatic speech translation processing capability of this software on NEC Electronics' MP211 (note 1*) application processor for mobile phones, at an operating frequency of 200MHz, proving that operation of interpretation applications is technologically feasible on small devices like mobile phones.
Supporting a 50,000-word rich vocabulary, this software realizes automatic speech-to-speech interpretation of travel conversation through the development of a new parallel speech recognition method (
note 2*) for single-chip processors with several CPU cores, and a compact, lexical-rule-based, machine translation engine that unites dictionaries with grammar (note 3*) that is operable on small devices.
The features of this software include:
(1) A parallel, large-vocabulary, continuous speech recognition engine, which is built with a database consisting of a wide-range of conversation sounds and words that enables accurate speech recognition of spoken words.
(2) A lexical-rule-based, machine translation engine, which achieves high-performance translation of spoken words utilizing dictionaries/grammar, compiled from a wide range of language knowledge data.
(3) An advanced wave-concatenative speech synthesis engine, which realizes high-performance reading through an advanced, wave-concatenative speech synthesis method based on a wide-range of speech data.
(4) A total integration module that controls collaborative operation of the speech recognition engine, the machine translation engine, and the speech synthesis engine realizing automatic translation on a single processor for mobile phones.
With the advancement of an information society and increased freedom of movement across borders, the dynamic development of technology supporting automatic speech interpretation and translation to support communication between different languages is rapidly progressing.
NEC's developments in this area include:
Automatic Japanese – English/English – Japanese translation software for notebook PCs in 1999
Commercial launch of "Tabitsu" (American English version),communication software supporting English travel conversation, in 2001
PDA-operational Japanese – English travel conversation, automatic speech translation software in 2002
The next natural development for NEC was to expand this technology to small, light-weight portable devices that can be used anytime, anywhere. However, in order to achieve this goal it was necessary to realize large CPU power, required for speech recognition, and machine translation technology for interpretation, which are both exceedingly difficult to achieve on low-power multi-core processors for small devices such as mobile phones. NEC has accomplished this development through the synthesis of its proprietary parallel speech recognition technology and its compact machine translation technology with its multi-core processor technology.
NEC will continue to advance research of its speech recognition and language processing technologies toward the realization of a society where communication is possible anytime, anywhere.
-------------------------------------------------
Note:
With the combination of three ARM926EJ-S CPU cores and NEC Electronics' digital signal processor, the single-chip MP211 processor offers high-performance parallel processing capabilities optimized for applications such as mobile phones that are sensitive to power consumption requirements.
Through adoption of an acoustic look-ahead technique that can reduce word-search space, NEC's proprietary speech recognition method realizes acceleration of the entire interpretation process, dividing it into three steps comprised of recognition processing, reconstruction and maintenance of accuracy.
This is a translation engine based on dictionary storing of each word's lexical rules that can easily expand both general translation rules and individual translation rules for fixed form expression, and which realizes excellent software downsizing and enhancement of translation quality.
