Microsoft Translator accelerates use of Neural Networks across its offerings

Exactly one year ago, Microsoft Translator made a new AI-powered technology available to everyone: neural machine translation (NMT) . Since then, the team has been working hard to expand the use of NMT to improve translations in the Microsoft Translator Text and Speech APIs, as well as in all of Microsoft Translator’s supported products.

Today, the Microsoft Translator team is announcing several developments in NMT technology, making advanced AI translations more accessible, no matter how you use them:

• 10 new languages are now available for NMT in both the API and the apps
• All of the API traffic for Chinese and Hindi is now powered by NMT, developers do not need to do anything to use it
• New “hybrid” translation is available for API and apps users alike, bringing benefits of neural translations to languages not yet available on NMT
• Speech translation is now powered end to end with LSTM technology with the release of speech recognition LSTM systems in the speech API
• NMT is now also available as an on premises service

NMT technology has transformed machine translation, providing major advances in translation quality over the existing industry-standard statistical machine translation (SMT) technology. NMT better captures the context of full sentences before translating them, providing much higher quality translation and more human-sounding output. Learn more about how NMT works on the Microsoft Translator website, and watch the video below to see how NMT is integrated into Microsoft Translator’s speech translation.

You can try out the new NMT powered translation today in Translator for Bing, the Microsoft Translator apps, Edge, the Translator PowerPoint add in, and of course, the Microsoft Translator Text and Speech APIs.

Introducing 10 New NMT-Powered Languages

Microsoft Translator adds 10 new languages to its list of NMT systems—now at 21 languages and growing! NMT now powers translations in the following languages (new languages in bold).

100% of Chinese and Hindi translations moved to NMT

Microsoft is making NTM more easily available to developers. Specifically, all Chinese and Hindi translations to and from English are now using our new NMT systems by default.  This means that developers do not need to call the “generalnn” category in their apps to benefit from these new systems. For the other languages this category is still needed, but additional languages will be switched to 100% NMT in the coming weeks and months.

For developers that want to keep using SMT systems for the time being, for instance if they have built customs systems using the Hub, the category “SMT” can be used when calling the API.

“Hybrid” Neural/Statistical translation

With hybrid translation, when only one of the two languages you are translating to or from is NMT-powered, you will still see an increase in translation quality.

Because of the scarcity of available training data between languages and the exponential complexity of building and running dedicated systems for each language pair, machine translation systems use English as a “pivot language” to translate from one language to the other. This means that each time you ask a machine translations system to translate from, say, Chinese to Spanish, the translation system usually translates Chinese to English, then translates the English to Spanish.

With hybrid translations, when at least one of the two languages has an available NMT system, Microsoft Translator will automatically use NMT for that section of the translation. This portion of the translation will improve, making the entire translation better. View our article on the Microsoft support forum to learn more about hybrid translation.

End to end LSTM neural network powered speech translation to dramatically improve speech translation quality

Speech recognition is moving to advanced LSTM neural network architecture. Combined with an increase in available speech data, LSTM speech recognition improves quality (measured by the industry standard “word error rate”) up to 29%, depending on the language. This has a direct impact on the quality of the machine translation, since the more accurate the speech recognition is, the more accurate the resulting translation will be.

Microsoft Translator’s NMT uses LSTM technology– speech translation is therefore now using  LSTM technology from end to end. Try out end to end LSTM speech translation capabilities with Microsoft Translator live feature in the Translator PowerPoint add in, on http://translate.it and the Microsoft Translator apps, or in the Translator Speech API.

On premises neural networks

For organizations that require additional data security, neural network translation is now also available as an on premises service, using the organization’s own servers rather than the Microsoft Azure cloud service. Learn more about Microsoft Translator’s on premises offering.

Learn more in the blogs from Microsoft Research and Cognitive Services.

The post Microsoft Translator accelerates use of Neural Networks across its offerings appeared first on Microsoft Translator Blog.