Excerpt: pages 15–27 — Translation Technology Categories – This report was written for a TAUS assignment.

Here we break down the existing translation technologies into categories based on three main principles:

• Recognizability – The categories are labelled according to main functionalities of the translation technologies;
• Target users – Some technologies are developed for the translation process and aimed at linguists and translators; others offer features that are mostly related to project management, and, therefore, are useful for language service providers;
• Usage – The majority of the technologies discussed in this report are meant for use by specialists; few others (for example, the technologies included in the Streetwise category) are aimed at the general public.

These categories are not set in stone. Just like most established software houses, translation technology providers tend to develop their products over time by adding new functionalities. It is not unusual to see how a technology that starts out as a basic computer-assisted translation tool or a translation management tool quickly evolves – through incremental developments and integration with third parties’ products – to become a comprehensive localization environment.

Technology as a mirror of the business environment?

We could make a bold statement and say that the development of new translation technologies is just a reflection of the fluidity of that “gig economy” that characterizes many industries, including the translation industry.

In 2009, when the effects of the financial crisis started to affect the most vulnerable categories of workers, Tina Brown, former editor of Vanity Fair and The New Yorker, brought the concept of “gig economy” to the general attention in an article in The Daily Beast. Brown’s article opened with what at first read like a depressing statement: “No one I know has a job anymore. They’ve got gigs.”

In Brown’s meaning, gigs were a bunch of free-floating projects, consultancies, part-time bits and pieces people stitched together to make a living. She used a slang term coined in the 1920s by jazz musicians and extended its meaning to the act of combining multiple skills, talents and abilities to generate a decent income.

Only in recent years, the term “gig economy” has become a Wikipedia entry, thus entering the public domain, even though it is no news for the language industry that has been applying the “gig model” for some three decades.

A key characteristic of the translation industry is its vast cross-company and freelancing spread, with projects starting at one buyer, branching out to different vendors (sometimes located in different countries), who split their portion of the project and subcontract it to several other smaller vendors.

These smaller vendors, in turn, split their assignment in even smaller chunks and hand them off to individual players (in the majority of cases these are freelance translators). The role of the majority of the players within a localization project ends up being a gig. The more skills you have and the more tools you can use, the more gigs you get.

Despite this technological and business fluidity, we think that the categorization we propose will help beginners navigate the complexity of the current translation technology landscape and sift through the different offerings to make informed decisions. For more experienced users and business decision-makers, the report shares insights on future trends. Investors and policymakers will benefit from analyses of underlying value propositions.

Methodology

Because of the creative busyness and the momentum of the translation technology industry, new and innovative technologies and tools are introduced every day. Unfortunately, not every single new breakthrough is meant to stay. Some may be barely viable; others may never come to fruition. Some are meant to solve thorny problems; others introduce new ways of working. Only time can tell which innovations are the ones that really matter.

For this reason, in order to provide a realistic overview of the current translation technology landscape, we turned to translation technology providers – both within and outside the TAUS community – with an active presence in the translation industry to collect information about their products. The technology profiles can be found from page 31 onwards.

Translation Memory Tools

A translation memory (TM) is a repository of previously created human translations to be reused. Typically, this happens by looking for matches between source-target language pairs stored in the TM and presenting them as translation candidates. In this respect, a TM is expected to boost productivity by increasing efficiency because translators may or must leverage – partially or totally – previously translated texts.

TMs have been around since the late 1980s but they became a significant commercial entity only in the late 1990s, when the ideas behind them – reusability and leveraging – gave way to a new business logic based on discounts and higher productivity.

In the last two decades, providers of TM tools have been further developing their products by incorporating features previously only available through distinct applications. Nowadays, most TM tools can support and facilitate the main steps of the translation process: from terminology mining and management for consistency and accuracy to multiple-format conversion for code, formatting, and layout protection; from word count to job statistics; from quality assurance (QA) to MT integration.

Productivity

Initially, TM tools were only addressing reusability as a way to improve consistency and sustainability. Over the years, technology providers have focused on other important aspects (i.e., manageability, adaptability and data exchange through multi-format capability): the TM tools currently available fit into the typical translation workflow and provide a big boost to translators’ efficiency.

Productivity is at the core of any TM tool. It is achieved by reusing and leveraging existing translations, i.e. comparing files against available memories and pre-translating them with exact matches. For documents and projects with a lot of repeating texts, this can yield significant savings in terms of costs and time.

Recently, with the introduction of new linguistic analysis capabilities, pre-translation can also be run through a combination of TM and suggestions from MT engines. These suggestions are combined with fuzzy matches, i.e. with TM entries scoring similar source texts for improvement.

Fuzzy matches are used as a style and memory aid for human translators to use any relevant piece as needed. Although not yielding the same cost savings, they are used to reduce errors and to encourage the use of a consistent writing style.

Translation management

Because of the increasing complexity of translation tasks, translation businesses and departments must now manage projects consisting of many files in different formats and many language combinations, using teams of two or more translators. The majority of TM tools currently available can provide project management functions that vary from basic to sophisticated, to help translation providers control overhead costs and budget planning.

Businesses running TM tools – whether desktop, server or cloud editions – typically process project files by quantifying the number and type of matches and repetitions in order to calculate translation costs and turnaround time. They can also monitor the progress of the various translation projects, whether they are assigned to a single linguist or to a translation team.

Large translation businesses with demanding clients look elsewhere for savings, massively exploiting ancillary functions like document alignment – to build TMs from legacy and third-party projects – or terminology mining features – to build terminological databases.

Client-server Model

The client–server model is an application structure to partition workloads between a service provider (the server) and many service requesters (the clients) over a communication network. The server hosts and runs one or more applications that share the resources with the clients.

Up to now the client-server model has been quite popular among translation businesses and departments looking for a way to centralize their resources. The typical implementation consists of a server – usually hosted on the customer premises – which, in turn, interacts with client software installed on employees’ and translators’ machines.

Main drawback

Despite the many advantages offered by a centralized architecture, there has always been a major issue: interoperability, i.e. the capability of a product or system to work with other products or systems. Interoperability implies exchanges between vendors, products, and product versions. The lack of fully applied standards in translation is the main reason for the request of higher interoperability.

Translation software generally uses exchange formats (TMX, CSV, etc.) to ensure interoperability, i.e. data exchangeability. Unfortunately, this often leads to loss of information, reducing interoperability only to mere compatibility. Users are increasingly hostile to lock-in with any software vendor, and have constantly been searching for ways to reduce their Total Cost of Ownership (TCO) while coping with the steeping demand for higher productivity.

Cloud-based Model

To offer a solution to the high IT/TCO costs associated with a client-server solution (i.e., server procurement and maintenance, licensing and updates on a large number of machines), some translation technology providers have developed cloud-based interfaces and/or on-demand licensing systems; others have re-tooled their client-server solutions as cloud-based products.

The popularity of cloud-based translation is due to the need on the translation vendors’ side to maintain full control over their TMs, terminology databases, vendor bases and project data as well as stringent security requirements sometimes imposed by clients.

The migration from a desktop or client-server version to a cloud-based version has become possible thanks to the ubiquity of web browsers and their convenience as clients, but also because of the option of a quick update of web applications without the necessity of distributing and installing software.

The most interesting features of web-based TM tools are the inherent support for cross-platform compatibility, the ability to support users on mobile devices, and the licensing model (per user or volume-based), which enables customers to upgrade and downgrade as needed.

Terminology Management Tools & Repositories

Terminology Tools

Terminology tools are used to compile, mine, register, store, organize, search, translate, and share terms and terminological databases (termbases) on a customer, project or asset level, in order to ensure that the correct terms are leveraged consistently according to a set of rules. A basic terminology tool is a searchable database that contains a list of approved terms, with rules and annotations regarding their usage.

Most terminology tools are integrated into TM tools and typically used in conjunction with TMs to pre-translate recurring words and phrases, and to avoid translating items that should be left as is (e.g., non-translatables such as brand names). Users can click on a suggested term to paste it directly into the editor.

Most TM tools today also provide basic terminological QA functionality with the goal to detect inconsistent terminology usage, catch common errors, and produce a standardized output. With integrated terminology tools, users can also search existing termbases and perform contextual lookups in source material and bilingual reference memories.

Terminology tools usually also consist of a mining tool to extract relevant terms from a large and structured set of texts (corpus) using a statistical (frequency-based) or linguistic (morpho-syntactic) approach, sometimes in conjunction with an existing termbase to detect new occurrences. These tools prompt the user with a list of plausible terms (candidates) to choose from. Term extraction can, for example, be applied to parallel corpora to create bilingual glossaries.

The same termbases used for translation can effectively be used with controlled authoring tools and vice versa. Termbases are not only useful to define how a set of words, phrases or proper names should be translated (or not) or to reduce translation costs, but also to encourage consistent vocabulary and style, and prevent dissonant translations of frequently recurring phrases.

Terminology Repositories

A termbase is a database storing terminological data. More and more often, with the convenience presented by the World Wide Web and the related technologies, larger organizations are consolidating their termbases and migrating them to central storage locations from which users may search the databases and retrieve any terminological information.

These databases are generally free of charge and only require a registration. They contain multilingual terminology data and are in most cases publicly accessible termbases belonging to international organizations. The organizations running these databases typically provide a simplified search user interface, and a staff-only access door for the manipulation of data stored in the database .

Some of the free-of-charge termbases have also been made available for download in exchange formats like CSV or TBX, to be used in conjunction with a terminology tool and to be integrated in a translation environment, as they cannot be searched directly from a translation editor.

Alongside with termbases, the Web offers many other repositories following a different approach. Some of these are not properly terminological applications, but more similar to multilingual knowledge bases and encyclopedias.

Machine Translation Platforms

MT may be considered the grandfather of all translation automation technologies. And with good reason. From its early days MT has been thoroughly investigated. Tons of paper and ink have been spent to document the ongoing research and the repeated efforts to increase extensive use of MT.

Only in recent years, though, MT has been widely adopted; even the translation industry has been trying to come to terms with this new technology that, through numerous implementations, has overcome hostility and skepticism.

With the introduction of statistical engines, MT has made a dramatic progress in terms of speed and quality, and it is now used in the context of professional translation. A good example is MT engines integrated within TM tools offering ‘suggestions’ for what is now called ‘interactive post-editing’.

The booming of free online services has turned MT into a viable mainstream technology. It has also been creating a paradigm shift in the demand for translation: today MT is widely leveraged for content once doomed to oblivion because of the high cost of human translation. The so-called zero-translation option is now a thing of the past.

Rule-based machine translation

In its early days (some 70 years ago), the MT platforms available were using the rule-based approach. Rule-based machine translation (RbMT) has lasted until the present day, although the existing implementations are now being replaced. The new implementations, on the other hand, are limited in scope and restricted to minority languages for which little and poor language data is available but where the academic effort is greater.

RbMT engines work by parsing a source sentence, analyzing its structure, converting it into an intermediate construction, and, finally, transforming it into the target language. The advantage of RbMT engines is that they can be put to work almost immediately after installation.

The disadvantage is that each language pair requires a specific parser and as many dictionaries as the subject fields of work. While building glossaries could be a task for almost everyone, building or refining a parser with rules is quite complicated. For this reason, RbMT has largely been replaced by statistical machine translation (SMT).

Statistical machine translation

An SMT engine works by detecting patterns in reference monolingual corpora and looking for statistical correlations between source texts and translations in bilingual text corpora. An SMT engine does not leverage any grammar rules, but it generates probability scores for a given source text. It does so, first, by mapping it onto a reference source language corpus and a translation, and, then, by mapping the candidate translation onto a reference target language corpus.

The key advantage of SMT technology is that a translation engine can work with many different language pairs, and only large monolingual and bilingual corpora are required to train it for any language pair. The main drawback is that SMT engines perform well only when they are fed source texts pertaining to the same field as that of the training corpora.

Large volumes of clean and consistent training data are also needed. Even with suitable training data, SMT engines do not generally produce publication-ready output and require post-editing.

However, the requirement for training data opens the door to the customization of SMT engines for specific domains. Recently, the time necessary for customization has been significantly reduced, and most technology providers already offer SaaS-based SMT solutions so that users can create specific SMT engines in few hours. Also, a new and innovative technology called Adaptive Interactive Dynamic Learning MT is becoming more and more popular because it allows users to tune and adjust the engine dynamically in real-time.

Apart from a few semi-commercial solutions based on the free open-source Moses engine, most SMT platforms are cloud-based. They are fully customizable, mostly through a user-friendly web-based interface.

Although the licensing model is generally on-demand, most technology companies offer on-premise internal-cloud or private-cloud installations. The free online generic engines are not customizable. All implementations can generally be integrated – through APIs – into desktop or cloud-based TM tools.

Nowadays we are standing on the verge of even more exciting moment in the MT history: deep learning (DL) is taking MT towards much higher accuracy and finally brings human-like semantics to the translation process. More about this in the Trends section of this report.

Post-editing Tools

Post-editing is the adjustment done by humans of machine-generated translations in order to achieve a final product that fulfils agreed specifications, i.e. a level of quality negotiated in advance between the client and the service provider.

The improvements in MT quality and the large availability of platforms have led to an unprecedented demand for translation, making post-editing a common practice that allows for saving time and money. Post-editing is also important to track and understand issues in the performance of MT engines and to make the necessary adjustments to improve them.

Given the specific and different skills required, translators have been the preferred post-editors. Also, many researchers and teachers at vocational faculties have been working on this topic, with specific tools being developed to support instruction.

TM tools can also be used for post-editing. Right now they represent the best solution to facilitate the integration of MT into the typical human translation workflow. They do have three main limitations though: cost (most are proprietary tools only available as part of a major product distribution), limited flexibility, and lack of detailed statistics from post-editing jobs.

Therefore, dedicated post-editing tools are being developed following a user-centric design to reduce the slope of the learning curve and increase user efficiency. These tools follow the same approach as professional translation environment tools: they segment texts in distinct chunks and display the source text aside or above the corresponding MT output, which can be edited in a separate window. Unlike the professional TM tools, specific post-editing tools can handle only XML-based formats, typically a bilingual format like TMX.

Translation Management Systems

Translation management systems (TMS) are the natural evolution of translation servers as they enable translation businesses to monitor projects and control every single task according to a predefined workflow. The implementation of a TMS contributes to eliminate all manual tasks, thus cutting overhead costs and increasing project execution speed.

As outlined in the previous pages, nowadays many state-of-the-art desktop TM tools provide some basic translation management functions, but most TMSs are web-based and rendered with the Software as a Service (SaaS) licensing model.

The major advantages of the SaaS model are: • Accessibility from anywhere, through an Internet connection; • No local server installation required; • Pay-per-use or subscription-based payment methods; • Scalability; • Reliability; • System maintenance included in service.

The SaaS pricing model helps to bypass wasted time, money, and product limitations on traditional business applications, achieving a more efficient, centrally controlled business, and allowing access to all applications from anywhere (as long as an Internet connection is available), without having to worry about servers, office space, power, etc.

Capital expenditure is also reduced because a translation business doesn’t have to purchase servers or full copies of software. Because no local installation is required, software can be deployed more quickly while human resources are now free to focus on core business activities.

There are few drawbacks: from security and data integrity issues to the increased revenue cost of paying for the use of the services; from the dependence on Internet connection availability to the lower flexibility and the higher customization costs. The SaaS model offers a real-time collaborative translation platform, with parallelized processes instead of the typical serial approach. Licensing is generally done via subscription, with different payment plans for linguists and companies of different sizes.

All TMSs are developed around some common basic workflow management functions, consisting of deadline management and some mechanisms for translation resource management. The features that are offered depend on the level of sophistication and may differ according to the availability of a supplier module, possible connectivity to a content management system, the potential for being networked and the access to the system for clients and suppliers. TMSs also offer the basic technological requirements for crowdsourcing projects.

TMSs with a deep connection to a content management system are usually called globalization management systems (GMS).

Globalization Management Systems

A GMS is typically provided through an API to allow subscribers to customize the system’s appearance (the user interface) and manage the workflow (e.g. assigning roles and deadlines, monitoring costs, workload and quality parameters within a project, and managing routing and distribution of content), while maintaining full control over every single step of the translation automation process, from file analysis to billing. Some systems also have versioning control functions to detect changes in the content to be translated and trigger dispatching.

Localization Project Management

Localization project management systems are translation industry verticals for business-management software. They usually consist of a suite of integrated applications to collect, store, manage and interpret data coming from different business activities (from planning and purchasing to service delivery).

Localization project management systems can be seen as the equivalent of ERP (Enterprise Resource Planning) systems in the manufacturing industry. They integrate all (or the most relevant) business processes of a translation business, comprise plugins for integration with partner applications in a Service-Oriented Architecture (SOA), and integrate translation into development processes.

With the help of a localization management system users can plan, develop resource estimates, monitor scheduling, allocate resources, control costs and, in some cases, manage quality and run administration tasks. Localization management systems were initially developed for software localization following the client-server model, but nowadays they are extending to web-based services that enable customers to upload their assets and monitor localization progress.

Most localization management systems today offer advanced process automation capabilities covering projects planning, staffing and monitoring. Functionalities span from quote and order management to graphic workflow design and automation. Some systems also offer customizable procedures for automated vendor selection in order to optimize and speed up job assignment.

Proxy-based Localization Management

Proxy-based localization management platforms exploit the relay principle of computer networks known as proxying. A proxy is a computer system or application that acts as a substitute for another system or application. Specifically, proxy servers act as intermediaries between clients requesting a resource and the originating provider.

Proxy-based localization management platforms are built around proxy servers that receive traffic from global audiences who are routed – through these proxy servers – to source websites. As visitors browse a proxied site, requests go back to the source site where resources are rendered. The original language content in the response is replaced by translated content as it passes back through the proxy.

Localization proxy servers are typical reverse proxies that appear to clients as being ordinary servers: responses are returned as if they came directly from the originating servers.

Building a global website and having it properly localized in all the desired locales takes planning, skill and investment, which is hardly affordable, especially when reaching a global audience is a must. Retrofitting could be even more challenging. Proxy-based localization technology, therefore, offers a solution to many typical internationalization issues requiring heavy re-coding to support multiple languages, especially for e-commerce websites.

The management component is made up of an integrated or built-in translation management system, where all textual content of a website is extracted by the localization proxy servers and channeled through for translation.

The translation can be either MT or human translation, though in most cases, it is a form of computer-assisted human translation, resulting from a combination of MT, TMs and human translation. The localization management component is meant to put project management functionalities into the hand of localization buyers.

When users request the content, localization proxy servers provide the translation, saving the owner of the originating platform the hassle of localizing the content and storing all the localized versions – not to mention the nuisance of adding alternate languages down the road.

When properly configured as basically reverse proxy servers, localization proxy servers can hide the existence and characteristics of the origin servers, thus offering an additional security layer. They can also distribute the load, cache and optimize content, and rewrite URLs to match geographical or localization criteria.

App Localization Systems

Apps (applications) are computer programs that run on smartphones and tablets. Many mobile apps are increasingly standing in contrast to software applications running on computers, and to web applications.

App development requires specialized Integrated Development Environments (IDE), using emulators for testing. App localization systems are usually localization proxies with limited TMS capabilities. Their main feature consists of dedicated plugins (connectors) to connect to the original IDE or the Concurrent Versions System (CVS), i.e. the software development system that keeps track of all work and changes, and allows developing teams to collaborate.

Once integrated, the connectors automatically identify the translatable content within an app, upload it into the system, and constantly synchronize it with the app itself. Some app localization systems also prepare a localized app for release and submit it to the distribution store.

The most intriguing components in any app localization system are emulators, which provide pseudo-localization features to allow developers and localizers to deal with the intricacies of the limited space available on mobile device screens. Many app localization systems work in combination with external localization management platforms in order to provide functionalities like TM leveraging, progress monitoring, and workflow management.

Captive Translation Management Systems

Captive translation management systems can be either implementations of commercial SaaS TMSs to the exclusive use of a company’s clients (as part of a service agreement) or proprietary platforms specifically developed by a translation business to enhance process automation. In both cases, upgrades and improvements are made by the technology owner and can be limited in scope.

Customers looking for higher flexibility can pursue customized solutions that are not tied to a single vendor. Upgrades and improvements are made separately and in different stages by the various providers. These solutions require a much higher degree of integration, longer time for design and implementation as well as highly specialized dedicated staff.

Community Translation Platforms

User communities are invaluable sources of knowledge, because of their accumulated experience, their feedback, and their possible recommendations. Companies have learnt that engaging user communities and exploiting the huge amounts of information and energy present in them is essential for keeping their own products suitable and, at the same time, win customer loyalty.

User communities can represent an important factor, also in a translation/localization effort because of the valuable understanding of specific languages and cultures.

Most translators active in user communities are usually laymen. Although this does not mean they cannot produce excellent translations, a platform that enables them to run a collaborative effort is essential to get the best from their ‘experience’.

A community translation platform is a way to integrate open knowledge into a collaborative network following a social media model. The fundamental ingredients of a community translation platform are a role-based organizational model, a technological infrastructure for project management and translation, communication functionalities and, possibly, a voting mechanism.

In this perspective, community translation platforms are mostly translation management systems with an (online) friendly editor and collaboration features (chat, file sharing, forum, etc.) to invite and hire translators or delegate to contractors.

Community translation does not necessarily mean crowdsourcing translation, while a crowdsourcing translation/localization project strictly requires a community translation platform.

Middleware

With respect to translation software, middleware generally refers to a suite of connectors providing services for the integration of content management systems (CMSs) and TMSs by managing content flow from one system to the other. By using web services, middleware platforms establish direct application-to-application interactions, allowing users to programmatically create processes and workflows on top of open standards in order to exchange data between systems.

The integration is made easier with APIs, defining functionalities that are independent of their respective implementations. APIs are possibly the most common method for translation customers to integrate their own system with their vendor’s system. APIs are also used to implement localization proxies via remote invocations of services on common protocols.

In addition, since protocols are usually shared between different technologies, middleware platforms are increasingly built on APIs to exchange information, used as abstraction layers between different environments, with the aid of a language-neutral message format.

Controlled Authoring Tools

The cost of information is related to its clarity, conciseness and effectiveness, as well as to its reusability for different media, purposes, and audiences. To achieve all of this during content development, a possible solution is applying a set of predefined rules (style, grammar, punctuation) and approved terminology. This process is commonly known as controlled authoring.

In recent years, the increasing demand for automation and, more specifically, for MT has led to an increased use of controlled authoring as a means to improve content and manage translation costs. In this respect, controlled authoring tools assist writers to check compliance of documents to predefined writing rules. These tools are also called controlled-language checkers and automated editing software.

Controlled authoring tools parse texts and bring style, grammar and terminology deviations to the writer’s attention. Some tools allow for customization, i.e. users can specify which writing rules and terminology restrictions should be applied.

The beneficial effects of a standardized way of writing supported by controlled authoring tools can be identified in improved safety and customer service from clearer communication, less recalls and liability claims, enhanced content management and considerable cost savings related to reduced time-to-market, lower word count, shorter translation turnaround time and, finally, reduced translation costs.

Quality Assurance Tools

Quality Assurance (QA) tools are software tools used to compare source and target segments of bilingual texts in order to identify and resolve formal errors in bilingual translation files and TMs.

QA tools detect formatting, consistency, grammar, punctuation, capitalization, spelling errors, target segments that are identical to source segments, untranslated segments, omissions and incorrect non-translatables in the target language. At the same time, these tools can check compliance with project glossaries.

Other typical QA functionalities include length comparison, repeated words, double spaces, unit conversion verification, quotation marks, brackets, and tags. In most tools checking routines can be created using regular expressions. All detected errors are included in a report.

QA tools have a number of intrinsic limitations, the most important being the high vulnerability to false positives. This is generally due to different grammatical rules of source and target languages. Also, QA tools expect the source text to be correct, while a mistake present in the source text could have been rectified in translation. Although they don’t replace human editors and proofreaders – and despite their limitations – QA tools can help save time, especially with larger volumes, as the typical checks they perform are too time-consuming, too annoying – or simply impossible – to be done manually.

QA tools are also an irreplaceable help for translators, no matter how experienced they might be. In fact, all modern translation environment tools include a variety of QA functions. They are also invaluable aids in preparing files for post-editing or preparing data for SMT. In some cases, QA tools can be used to assess the quality of a translation sample.

Audio-Video Captioning

Captioning and subtitling are both processes of displaying text derived from either a transcript or a screenplay of the dialog or commentary in films, television programs, video games, and other visual display, usually at the bottom of the screen.

The distinction between captioning and subtitling is subtle. Both are transcription or translation of the script, but while subtitles only refer to dialogues, captions provide other relevant or interpretive information to complement non-speech elements. Captions can be ‘closed’ or ‘open’. In the first case, they are not visible until activated by the viewer. In the second case, they are always visible to all viewers.

Captioning tools are specialized software programs to digitally manipulate videos by making each frame accessible. Text is usually added to a sequence by marking the two ends between which the subtitle should appear. These markers are usually based on timecode, the standard numeric sequence for time reference for editing, synchronization and identification.

There is a large number of readily available applications for video subtitling today that help subtitlers perform all pertinent tasks in front of a single screen. In addition to deciding the exact in- and out-time of each subtitle, they can monitor the reading speed and length of subtitles, and choose the positioning and color of text. Although still rather expensive, professional subtitling software is now within reach of many translators, also thanks to the emergence of many free and equally reliable tools.

An early feature in most subtitling software programs is the automatic detection of shot changes along a timeline, showing the video track and the shot change boundaries. More recently, audio level indicators have been introduced to show changes in soundtrack volume as well as detect and distinguish speech from music or background effects.

Subtitlers can therefore skip speechless scenes while identifying the timing of speech points. Professional subtitling software programs also benefit from speech alignment technology for synchronizing audio and soundtrack, and assigning a timecode.

The spread of open source and free subtitling software together with the sprawling of the pervasive and unstoppable web has given rise to fan subbing, the practice of subtitling foreign films or television programs by fans into a language other than the original.

This has further spurred the development of applications and platforms built for this specific purpose, making the process fast and easy. These subtitling platforms are very easy to use and allow participants to concentrate only on the linguistic task.

An additional thrust has come from internet giants exploiting speech recognition technology for automatic captioning of user-generated videos. However, in this instance the quality and accuracy varies according to language, voice clarity and the amount of background noise.

Translation Apps

We are now long past the mobile tipping point: there are roughly 4.6 billion mobile users in the world (61% of global population). And, according to the International Telecoms Union, 43,4% of the world’s population is connected to the Internet.

The above mentioned data shows the apparent reality of a global population on the move, leading – as a natural consequence – to the ever-growing need of multifarious communication. This obviously entails an increasing demand for translations to help mobile users around the world find their way in the maze of messages and landmarks in a foreign language.

So far, we have described professional translation tools that cannot address this phenomenon, apart from free online MT. In fact, since its launch in 2007, Google has brought MT to the masses, making it free, quick and easy, and therefore significantly contributing to the demolition of the language barriers. A decade later, Google Translate counts 500 million daily users. But it is not alone to in bringing MT to the masses. In addition to Google and other internet giants like Microsoft (USA), Baidu (China), and Yandex (Russia), there are numerous other corporate and governmental MT services available.

App stores offer hundreds of translation-related apps to translate signs and print-outs using barcodes or QR-codes, spoken text through megaphones (currently limited to Japanese into Chinese, Korean and English) or a mobile device microphone, or messages within the most common messaging apps.

Most of them stand on Google’s shoulders, thus requiring an Internet connection, but they do strive to distinguish themselves with some useful (or just cute) features. There are also wearable translation devices (again, currently limited to translation between English, Japanese and Chinese) shaped as a pendant or a small earpiece or otherwise.

There are few translation apps that are academic-oriented and government-funded to support teaching and archiving of the languages and cultures of indigenous peoples; those that do support this are the result of privately funded initiatives to translate news and social media content in order to verify real-time news anywhere in the world and help journalistic investigative work.

Speech-to-Speech Translation

The goal of speech-to-speech translation (S2S) is to enable instant oral cross-lingual communication between people not sharing a common language. S2S translation is the nearest thing to the universal translator devices often depicted in many science fiction books and movies. Think of the Babel Fish in The Hitchhiker’s Guide to the Galaxy by Douglas Adams, the “microsofts” in Neuromancer by William Gibson or the TARDIS, the time machine annex universal translator, used by Doctor Who.

In the last four years, internet giants have brought S2S translation to the real world with popular applications, mainly intended for mobile devices, like Google’s Voice Translator and Microsoft’s Skype Translator. This was made possible thanks to the combination of existing technologies that have become more accurate, i.e. speech recognition, MT, and speech synthesis.

A typical S2S translation system operates via a three-stage process: an Automatic Speech Recognition (ASR) unit converts spoken language into text; then an MT engine produces a translation of this text in the target language; finally, a Text-To-Speech (TTS) unit converts the resulting text into speech.

The speaker of source language speaks into a microphone and the ASR unit recognizes the utterance, compares the input with a phonological model – consisting of a corpus of speech data – and converts it into a string of words, using dictionary and grammar of the source language based on a corpus of text in the source language.

Synthesized speech is created by concatenating pieces of recorded speech that are stored in a database or on a model of the vocal tract and other human voice characteristics to create a completely ’synthetic’ voice output. Specific problems may occur due to incoherence and unclear word boundary of spoken language, although this generally presents less complex structures and less vocabulary. Also, in adverse environments, noise and distortion are both problematic for speech recognition and hence for MT.

A broadband internet connection is essential to overcome the critical challenge of storage and memory of smaller and smaller mobile devices that are supposed to house and process large corpora and complicated programs, including noise and distortion reduction algorithms.

This raises serious confidentiality issues, especially when S2S translation systems are envisaged to provide language support to the military during operations in foreign territories, or in a healthcare setting, where privacy issues could arise as everyone in the surrounding area could overhear any conversations.

A major improvement is expected from the unprecedented growth of data, starting from the already massive volume of audio data for ASR training. Another important improvement is expected to come from the implementation of neural networks for machine learning. Indeed, as we shall see in the Trends section of this report.