Sep 26, 2011 (11:09 AM EDT)
SAIC Takes On Google With Speech Translation Apps
Read the Original Article at InformationWeek
SAIC Linguistics provides both software and software-as-a-service technologies, as well as human translation services. Jonathan Litchman, senior VP of operations at SAIC, said the company has been providing translation solutions for years, but its latest offerings are based on the acquisition last November of machine translation specialist AppTek, which had a unique hybrid approach to statistical and rule-based machine translation technologies for both text and voice. "We've made a commitment to continuing to evolve the technology and are putting significant R&D funding and effort into continual enhancement of the core technology."
One of the ways SAIC plans to dramatize the power of the technology is with an iPhone app, to be released shortly, that will allow travelers to speak into their phone and have translated words and phrases come out the other end, in English, Spanish, or Arabic. To prevent voice recognition errors from complicating the translation, the app will first display a text rendition on screen to allow the users to make corrections before the translation software does its work, Lichtman said. All this will be powered by a cloud service, although the app will also have an offline mode that allows it to play translated statements prepared in advance, he said.
"We certainly hope to make money on the consumer app, but that is clearly not our focus," Lichtman said. The bigger purpose of releasing this to the market will be to expose the power of the translation technology to more people than would be possible through an enterprise sales effort alone, he said.
[Will IT managers soon be running their largest servers in the data center from a smartphone? See Run Your Data Center From An iPhone.]
Another mobile translation application that SAIC just fielded for public health officials in New Jersey is an English-Spanish translation application for the iPad, focused initially on education about influenza vaccinations. "In this case, the software is completely local to the iPad--we're not relying on a 3G or Wi-Fi connection," Lichtman said. The software can operate in speech-to-speech, speech-to-text, or text-to-text translation modes and is customized to the medical domain, meaning it can be adapted for uses beyond the influenza application, he said.
Translation technology is the subject of intense U.S. military research, because of the need to improve communication in operations like those in Iraq and Afghanistan. Automated translation can also be applied to commercial applications, such as chat translation for customer support.
SAIC's translation platform tries to combine the best of two schools of machine translation. Statistical machine translation, the approach used in products such as Google Translate, works by analyzing large samples of previously translated documents to determine which words and phrases are probably equivalent. Given a large enough sample, statistical translation is the fastest way of generating a large vocabulary of translations that are probably accurate, most of the time. Rule-based translation is the more systematic, labor-intensive approach of encoding the grammar of a language and the meaning of words in software.
"The hybrid approach allows us to quickly develop new languages, accurately," said Hassan Sawaf, formerly one of the leaders of AppTek and now chief scientist for language technology for SAIC. The statistical approach works well for languages commonly used on the Internet, he said, but takes a lot of data and processing time to work. When targeting one of the world's more rarely spoken languages, SAIC can instead put more emphasis on manually defined rules "so we don't need as much data," he said.
Lichtman said translations between languages and between voice and text are part of the same technology platform in SAIC's solution, whereas competitors that try to do direct translation of speech are typically integrating with another vendor's speech recognition and text-to-speech software.
Sawaf said speech-to-speech translation is still a multistep process because first speech is translated to text, then the text is translated, and finally voice is synthesized for the translated text. "Speech to speech is still bleeding edge," he acknowledged. But with the right understanding of the application domain (which narrows the vocabulary of words likely to be used), it's starting to find practical uses, he said.
SAIC produces solutions that produce more accurate results because they are customized to a specific knowledge domain, like medicine or consumer travel. Contracts range from "the low tens of thousands" to "over a million," based on the scope or complexity of the customer's needs, Lichtman said.
Automation and orchestration technologies can make IT more efficient and better able to serve the business by streamlining common tasks and speeding service delivery. In this report, we outline the potential snags and share strategies and best practices to ensure successful implementation. Download our report here. (Free registration required.)