In February, I generated a few comments from both suppliers and the general public when I posted this note, which some readers saw as critical of Google’s Web-based efforts to support so-called speech-to-speech translation. My point at the time was to praise Google; not so much for making the effort to bring highly-accurate and reliable machine translation services to the Web “in a matter of years” but for using its Web site to make machine translation (MT) services that work adequately for an impressive number of language pairs and is accessible to anyone with access to the Web.

In recent months, we’ve seen the community of companies offering MT taking significant steps forward. Perhaps lured by market assessments that valued global spending on “language translation and education services” at $12 billion in 2008. Realistically, that may represent a market peak, as relatively inexpensive machine translation solutions gain acceptance in a growing number of use cases where they can easily replace more expensive “human translators”.

In addition to Google, we’ve seen a couple of impressive new services launched in the last couple of weeks. One worth noting is the Trippo™ VoiceMagix application for Apple’s iPhone. Finnish mobile application specialist Cellictica developed the app and made it through the Apple’s vetting process to introduce it through the iTune’s AppStore a few weeks ago. This press release describes the application and also includes a link to a short video that illustrates its functions. Cellictica participated in the Nuance Mobile Developer Program and has successfully integrated both Dragon Dictation for speech-to-text transcription of the utterances to be translated and Nuance Vocalizer for text-to-speech rendering of the translated phrases.

The VoiceMagix app supports spoken input in English (but text input of a total of 26 languages). It performs speech-to-speech machine translation from English into fourteen different languages, including Chinese, Dutch, French, German, Greek, Hindi, Italian, Japanese, Polish, Portuguese, Russian, Spanish, and Thai. For the other 13, it renders written results in the native characters or script. Because of my own limited range in terms of languages spoken, I can’t vouch for the accuracy of the app in many of the supported languages, but the friends and family members who are native speakers in French, Hebrew and Spanish were duly impressed with the iPhone app. According to Cellictica’s press release, “Trippo VoiceMagix also runs Android™, Windows Mobile®, Java (J2ME) and BlackBerry®, and the company is planning to make the app available to other handsets through major app stores “in the near future.”

Cellictica is taking a decidedly mobile approach to MT. At the higher end of the spectrum a Florida-based company called LinguaSys is noteworthy for its ability to expand MT’s wingspan by accelerating the amount of time it takes to bring additional “language pairs” in support of a broader set of multi-lingual use cases. LinguSys offers what it calls “language middleware” and its proprietary Carabao MT Engine to shorten the time it takes to integrate natural language translation of what CEO Brian Garr calls “short shelf life” interactions, a term he uses to define text chat, e-mail, web pages and documents that require rapid translation in support of business objectives.

Garr told us in a recent interview that the secret sauce required to expedite development of highly-accurate MT between two new languages is the use of a “hybrid” approach. As a 20+ year veteran in the MT community (having served as CTO at one of the first MT specialists, Globalink, in the 1990s) Garr had observed the long-standing schism between solutions that use “statistical language modeling” (SLM) versus the ever-popular Hidden Markhov Models (HMMs) or “rules based” models to accomplish accurate machine translation. As Garr put it, “You were either a stats guy or a rules guy.”

LinguaSys’ Carabao engine provides highly accurate results by using both approaches. (He calls it “hybrid”. I’d call it Recombinant). Because the statistical approach doesn’t care about the “meaning” of what is said, it merely needs a large enough database or “corpus” of matched utterances to build its statistical model and make a good “first guess” at a translation. As it gains experience (and its “corpus” grows) the accuracy improves. essentially getting a statistics-based model to “guess” at the translation (without tackling “meaning”) and then applying rules to confirm the accuracy of the original rendering.

As for deployment architectures, LinguaSys developed TransGen, which is its User Interface platform which enables Carabao to be put to use as part of Web services that support translation of documents, chat, email or other text input into Web sites. For mobile users, LinguaSys has also developed apps for iPhones and Android-based devices. This, in many ways, puts it in direct competition with Cellictica, for the much narrower speech-to-speech rendering market. By contrast, LinguaSys’ approach conforms to the rules of a successful RC implementation where it is instantiated as standards-conformant middleware that can be integrated (or “mashed up”) with a large company’s existing workflows and business processes.

In conclusion, we think that MT has crossed a critical threshold in market acceptance. Prospective users understand what it does and companies like LinguaSys, Cellictica and Google are bringing solutions to market that work sufficiently well enough to build trust among users. Opus Research is adding MT to the short list of catalytic technologies that accelerate deployment of Recombinant Communications solutions and promote more efficient network-based communications and commerce.

Categories: Articles

Tags: Google, Mobile Speech Apps, Recombinant Communications, Machine Translation