Report by Yulia Akhulkova. Interactive tracker by Aleksey Schipaсk
Last updated: September 1, 2024
Introduction to the Language Technology Radar
Since 2018, Nimdzi has been tracking the language technology space. In 2024, the language technology market is more diverse in its actors and complex in structures than ever before. And while our yearly report, the Nimdzi Language Technology Atlas, has been serving the language industry (and beyond) well, offering a unified view of the modern language technology landscape and insights into major technological advancements, it has been a snapshot of the landscape of tools at a given point in time.
As this landscape has grown to over a thousand products, we made a decision to support our tracker with interactivity that this dynamic market deserves. What we have created this year is a curated catalog of language technology companies. It’s no longer just a snapshot, it’s now a constantly updated database of products that brings visibility and transparency to the language technology market and helps with related decision-making.
Technology providers are welcome to use the Nimdzi Language Technology Radar both to benchmark their competition as well as to find partners. Investors can refer to it to gain a better understanding of the leading market players. Linguists and buyers of language services can see what tools are out there to help them in their day-to-day jobs. Students of language programs are invited to check out the Radar to discover just how many tools may be a mere click away for use in their future careers.
Nimdzi’s new online catalog also shows that categories of language technologies overlap as they are used both as standalone tools and as building blocks in compound language technologies. Moreover, while some companies are fully focused on developing language technologies, others bundle professional services such as data, customization, and deployment, as well as additional services. So let’s dive straight into this diverse mosaic of companies and their respective solutions.
Language technology market backdrop
Taking into consideration how the language technology space is proliferated with many free-of-charge solutions (from machine translation to transcription and AI chatbots), it may seem that the language problem is finally solved. Far from it; as we reported in our 2024 Nimdzi 100, the language services industry keeps growing not despite of, but due to the fact that language technologies are continuously improving.
"While the vast majority of words are already translated by machines, automated captions and subtitles have become a commonplace on YouTube, and copilots are being deployed in various departments for productivity gains, human expertise and oversight is still a must for high-value, error-sensitive tasks. This, in combination of the explosion of content in our attention economy, is what drives the growth of the language industry, and the need for continuous innovation in the language technology space."
Laszlo K. Varga
Indeed, the number of free and paid products we tracked, checked, demoed, and studied for the Radar shows that there is a large market for language technologies, and it is growing. Language technologies are increasingly used as productivity enhancers in language services, as standalone automated solutions for non-mission critical language tasks, and as gateways to multilingual communication with or without human supervision.
Undoubtedly, generative AI (GenAI) solutions have made a spectacular entry into the language technology space over the past two years. From text-based large language models (LLMs) to multimodal platforms that can process text, image, and audio they present a new wave of potential in the industry, enhancing communications, trade, and productivity.
"As LLMs and GenAI solutions multiply, the real challenge isn't whether to use them, but which one to choose. The paradox of choice in today's AI landscape isn't about abundance, but about discernment. Too many options can overwhelm us, making it harder to find the right fit for our specific needs."
Renato Beninatto
Methodology
In 2024, we collected data from providers of over a thousand technology solutions. The data gathering has been based on four main sources:
- Ongoing research around new additions as well as changes to the companies in the Nimdzi Language Technology Atlas, including publicly available data (company websites, press releases, webinars, published research papers, etc.).
- Data submitted by the language technology providers themselves via a form, and then verified by Nimdzi researchers.
- Briefings and meetings with language technology companies held during the first 7 months of 2024.
- Professional experience of Nimdzi team members who regularly use and evaluate various language tools for end-user engagements.
These sources have given us a comprehensive understanding of the state of technology development in the language industry, which we are presenting in the new Language Technology Radar.
Categories of the Nimdzi Radar
Let’s start our journey by reviewing the definitions of key language technologies.
Interpreting systems
At Nimdzi, we coined the umbrella term ‘virtual interpreting technology’ (VIT) to describe any kind of technology that is used to deliver or facilitate interpreting services in the virtual realm. There are three ways in which virtual interpreting can be performed or delivered: via over-the-phone interpreting (OPI), video remote interpreting (VRI), or remote simultaneous interpreting (RSI).
As the name OPI suggests, two or more speakers and an interpreter use a phone to communicate. This is an audio-only solution and the interpretation is performed consecutively. VRI is also performed consecutively. However, in this case, there is both an audio and a video feed. Depending on the VRI solution, users and interpreters either connect via an online platform with video calling capability or via a mobile app. As for RSI, it directly evolved out of the field of conference interpreting and is intended for large online meetings and events with participants from many different language backgrounds. The interpretation is performed simultaneously — at the same time as the speakers give their speeches.
Interpreter management and scheduling (IMS) systems are also included in our definition of VIT because, even though they do not focus on delivering interpreting services, they facilitate them. An IMS is a useful tool that allows for efficient management of interpreter bookings for both onsite and virtual interpreting assignments. We have included machine interpreting (MI) solutions in our definition of VIT and are subsequently listing them in our tracker as well. "AI-powered interpreting", or Machine Interpreting (MI), is the transmission of a spoken message in one language into a spoken message in a different language using AI without the input of a human interpreter. With MI, interactions between people who speak different languages can be facilitated solely by technology. The final product is a synthetic voice producing the speaker’s message in a different language from the original.
Speech recognition solutions
Also known as automatic speech recognition (ASR) or speech-to-text (STT) tools. The section features solutions that focus on automatic transcription as well as automatic captions and subtitles. Many of the solutions listed here provide both options. However, as there is not 100% match between these two groups, we subdivide this category into two subcategories.
Audiovisual translation tools
Here, we feature various tools and platforms for audiovisual translation enablement: from multimedia localization project and asset management tools to AI-enhanced dubbing tools.
There are currently five subcategories in this category of solutions:
- AI-enhanced dubbing tools
- Project and asset management
- Subtitling editors
- Dubbing editors
- Remote recording
Integrators
Here we list systems that integrate other, third party, systems with each other. The middleware subsection discusses major companies that specialize in integrating various language technologies. The products in the MT Aggregators subsection not only provide smart access to MT engines, but support certain procedures around MT so that users can leverage MT in the best way possible.
Translation management systems
Translation management systems (TMS) are systems that feature both translation (and editing) environments and project management modules. Core components of a typical TMS include:
- Bilingual translation environment (source and target)
- Translation memory (TM)
- Termbase (TB)
- Machine translation (MT) (optional)
- Project management features
- Built-in quality assurance (QA)
You can check out many more features of a modern TMS using Nimdzi’s free TMS Feature Explorer. There are now many AI-related features!
Within the TMS category there are four main subcategories:
- Localization for developers. This subcategory is reserved mostly for developer-oriented tools that focus on enabling software teams through modern localization solutions. Whether it is an open-source TMS with basic functionality or a full-fledged cloud TMS that serves dev teams and enables the full software localization cycle, both fall into this subcategory.
- Enterprise-only TMS. Solutions oriented mostly toward enterprise-level customers. Can a freelancer buy a license to such a TMS? Not necessarily. In a more usual scenario, an enterprise may grant the corresponding licenses to the team.
- Proxy & JS-based website loc and updates on air. TMS solutions oriented toward web localization fall into this subcategory. Solutions such as a translation proxy or an anti-proxy approach both fall into this subcategory.
- Generic TMS for every customer profile. Here, some of the brands that are present in the three previous categories also appear. This is a subcategory for TMS-type tools that can be (and are) used by all user types: freelancers, agencies, enterprises. One for all, all for one, as they say!
Translation business management systems
Unlike TMS, translation business management systems do not have a bilingual translation environment, but focus on management features for translation project enablement. We call such technology a BMS or (T)BMS, since that’s exactly what it does: it helps manage business operations around translation.
Marketplaces and platforms
In this section, we feature platforms and marketplaces focused specifically on translation, interpretation, voice, and localization talent. In a marketplace, you can post a job and accept responses from linguists and other professionals who are interested in doing the work for you. Then you book this talent or directly assign the job to the chosen talent within the platform. If you’re a linguist, you sign up and set up your profile in the system, get vetted and/or tested (on some marketplaces), and then start receiving job offers. There is also the platform language service provider (LSP) option where you not only get access to a library of linguistic resources and agencies, but also to the workflows for the projects along with PMs who support you. You can upload your files to the platform, get an instant quote, and after quote approval and project completion, receive the localized files.
Machine translation
The section discusses major machine translation (MT) engine brands subdivided into four subcategories based on the MT providers’ specialization:
- Ready to use engines (Generic Platform/API)
- Custom/Trainable engines
- Niche/Legacy engines
- Standalone Machine Translation Quality Estimation (MTQE)
Quality management
This section is devoted to quality management tools in translation. It features three separate subcategories which correspond to three main product types in this area: QA tools, review and evaluation tools, and terminology management tools.
Multilingual Creation Tools
We have divided this category into two subcategories for now:
- Multilingual content generators. These are applications for generating content from briefs, templates, style guides.
- Writing assistants and predictive text input. These applications are for assisting with content writing via grammar and syntax corrections, tone and styling options, and alternative phrasing suggestions.
Large Language Models
Large Language Models (LLMs) is a new category that wasn’t featured before, but which we simply had to introduce in 2024. We’ll discuss this new category in detail further in the report.
Here we list:
- Generic LLMs: Large. Pre-trained, generic, foundational LLMs with more than 10 billion parameters (such as GPT-4, Claude, Gemini, Command R, Granite, Titan).
- Generic LLMs: Small. Even though the marker “small” in the category name sounds a bit controversial, we needed to make this distinction from the “large” group of LLMs. Here we list pre-trained, generic, foundational LLMs with less than 10 billion parameters (such as Phi3, Gemma, Llama-8B, Mistral 7B, Qwen 7B).
- Language-specific LLMs. LLMs fine-tuned and trained to support a specific language or language groups (often built on top of third-party open LLMs).
- Localization-specific LLM tools. Mostly LLMs that are fine-tuned and trained for localization tasks such as translation, quality assurance, automatic post-editing.
New category on the radar: LLMs
LLMs are challenging the way we categorize language technologies. Unlike any other tool on our radar, LLMs are by nature general-purpose machines. Depending on their pre-training, they can be useful in translation and localization jobs, other NLP tasks such as summarization, or even software coding. They present a new way of using language technologies, not just because they are general-purpose, but because they can be fine-tuned for specific purposes. There are LLMs that are fine-tuned for translation and other translation-related tasks or created to support a specific set of languages.
This general-purpose nature of LLMs, their rapid proliferation, and ease of access practically democratized language technologies. This resulted in a plethora of experiments with these new GenAI tools, from language technology providers and tech-enabled language service providers to practically any tech-savvy company. While experiments and proofs-of-concept seem easy to create, LLMs’ large scale, enterprise-grade deployments haven’t arrived yet. Most players are still trying to figure out how LLMs fit into their workflows and technology stacks. However, while buyers recognize the opportunity presented by LLMs, they don't necessarily have the capability as such, because dealing with language and language data is different from traditional approaches in the industry. This is not dissimilar to how neural machine translation was disseminated within the language industry and beyond.
C-level executives on the buyer side see LLMs and generative AI as a new wave of productivity enhancers, slowly realizing that MT already introduced AI into language tasks in 2017. Buyer-side language programs and LSPs now have to leverage their expertise and experience with language AI for the benefit of their wider organizations. We are experiencing a new wave of fundamentally novel tools that can help with a variety of tasks, and the language problem is just one of them. That is why we have dedicated a separate category for LLMs in the Nimdzi Language Technology Radar.
"Within less than two years, LLM-powered AI Assistants have gone from being a wonder and cutting-edge innovation to becoming the new standard. Today, it’s almost impossible to find a major software platform without an AI Assistant that allows users to perform tasks using natural language."
Nadezda Jakubkova
Still, at the time of this report, there are not a lot of LLMs specifically geared for translation or translation-related work. TowerLLM from Unbabel, mastering a number of translation-oriented tasks, including grammatical error correction, MT, and evaluation, is one example. DeepL announced they also have an LLM that they use for translation, LILT also uses LLMs for translation work, and Translated is developing their Lara LLM solution for translation.
"LLMs have quickly become a feature in language industry workflows and tools. We predict that the next year or so will be the time of productization of LLM-based language industry solutions."
Laszlo K. Varga
We expect to see that next to the very large models such as Claude, GPT, or Gemini, smaller and more efficient LLMs will emerge targeting specific purposes in the language industry. There will be a plethora of LLM applications for proven use cases such as automated QA or post-editing. At the same time, new, previously not feasible challenges will find their automations via the application of the new generation of AI platforms.
Some LLMs are already proving to be useful in language tasks, but they require very specific engineering; not just prompt engineering, but orchestration, and even some additional natural language processing work. Nevertheless, there are use cases where LLMs are already handy: terminology extraction, paraphrasing, translation style changing, and more. In addition, their ability to use context in language tasks is unparalleled by previous technology solutions.
"LLMs will soon become the new standard for MT, replacing traditional seq2seq models, while fine-tuning multilingual models for translation-specific tasks is growing into a key area of emerging competition."
Jourik Ciesielski
Changes and challenges in the language technology market
Are humans still in-the-loop of localizing text and speech?
Let’s face it: the human-in-the-loop paradigm has overtaken the human-only approach in the language industry. In 2024, the vast majority of language work is already being facilitated by technologies helping human language talent (linguists, interpreters, transcribers, voice actors, etc.). Productivity-enhancing use cases include information retrieval, question answering, summarization, content drafting, creation and editing, sentiment and intent analysis, and multilingual customer support.
Some language technologies have reached a maturity level where they can perform tasks automatically without human intervention — at least in some use cases, domains, and languages. Speaking of which, there is a strong asymmetry in the performance of language technologies across languages. We’ll look closely at this further in the report.
Orchestration
One area of the industry with almost fully automated workflows is translation management. And here we mean not only translation business management systems (TBMS). The buzzword now is orchestration, with the focus shifting towards end-to-end integrations of the different applications in the localization tech stack.
"In the evolving landscape of localization, the need for fully integrated processes and technology is paramount, as standalone solutions such as traditional TBMS can no longer meet the efficiency and scalability demands of global content operations."
Roman Civin
Third-party middleware providers like BeLazy and Blackbird.io have recognized this need and provided refreshing alternatives to the bulkier products in TMS connector pools. Phrase has followed suit with the launch of Phrase Orchestrator.
There are also enterprises that don’t use TBMS tools but leverage integration solutions to manage their whole ‘business of translations’. They usually rely on software outside of the language technology space, namely tools like Excel, Google Sheets, Notion, ClickUp, Asana, Slack — all integrated with the help of robust middleware solutions that provide the process orchestration, automating everything possible for humans to only set up and finetune the workflows, and then leverage the automation perks.
However, in terms of the actual work that needs to be done in these orchestrated workflows, humans are for now indispensable for understanding linguistic nuance and cultural sensitivity. That is why in the vast majority of professional localization use cases, humans remain in the loop regardless of the advancements in automation or AI, particularly in neural machine translation (NMT) and LLMs. Major players in the language technology arena are trying to combine the best of AI with human expertise in a way that will deliver quality at scale.
Audiovisual translation and AI
The “human vs machine” question is relevant not only for localization management and written translation. Another area undergoing the AI boom and respective concerns about replacing human talent is audiovisual translation. Notes on the “first movie fully dubbed with AI” have already appeared in the news. For example, earlier this year, Camb.AI and Vox Distribution announced the “world's first movie dubbed with AI” with the release of the film “Three” (originally produced in English and Arabic) in Mandarin Chinese.
Big tech companies are also constantly announcing breakthroughs in AI(-powered) dubbing. Сompanies such as Microsoft, Google, Amazon, Nvidia, Baidu, ByteDance, and Alibaba, all offer speech-to-speech translation (S2ST) capabilities. But in the majority of cases, S2ST still uses a cascade architecture, which we already described in our last year’s report: automatic speech recognition/speech-to-text, then machine translation, and finally, text-to-speech/synthetic voices. And while developers of such solutions put a lot of focus on how the different modules in the cascade interact, humans are needed to continuously finetune the pipeline to improve results.
A view of how speech-to-speech translation is created in a cascading model.
Components on the bottom row are also used in other compositions or in standalone applications.
Orchestrated together, they create a novel technology.
Indeed, combining modules in a cascade does not necessarily result in a functional system. It works well enough for limited use cases, such as podcasts, e-learning courses, or corporate presentation videos. In multi-speaker settings, however, additional components are needed, such as speaker recognition, segmentation, etc.
For AI dubbing of multimedia content, one would also need script adaptation, voice acting, audio mixing, and more. This is less of an issue in non-live S2ST, where a pre-recorded audio or video gets processed and there is space for manual intervention and correction — with humans staying very much “in the loop” for a proper result when implementing this technology.
At the same time, Meta and Google are developing a direct method for S2ST solutions, which sidesteps the cascade using spectrograms instead of text-based operations. These multilingual models are not for commercial use but indicate a future direction of S2ST. Big tech’s platforms also offer an opportunity for smaller companies to compile S2ST solutions and productize them, especially those that are in the TTS and AI dubbing space. Examples of this include companies such as ElevenLabs, Deepdub, and Resemble.ai.
Automatic captions and subtitles
Another language technology “in the AI question” is subtitling. Here, it is important to differentiate automatic transcription and captions from subtitling. For now, automatic speech recognition (ASR) and particularly transcription have become commodities, especially with the help of Whisper and similar solutions.
But while the quality of ASR for clear source audio content is acceptable, especially for high-resource languages, low-resource languages do not have enough data for model training and attention from major tech developers, which is restricting the advancement of quality for these solutions.
A plethora of apps already offer automatic (multilingual) captioning with various quality outputs for individual use cases or as a feature of popular social apps by big tech companies. Automated captioning has also been made available on major streaming platforms, and while its quality is not perfect, it is becoming more accepted and used despite its current flaws.
At the same time, multilingual subtitling volumes have exploded, as it helps to reach the widest possible audience with less effort than dubbing. But subtitling a large multimedia piece for a global audience is not as easy as automatic captioning of an influencer’s blog. Human specialists therefore still have a job to do here.
How TMS providers implement AI and LLMs
Now, let’s see how TMS providers are implementing AI and LLM-based features into their product portfolio. One of the most observable trends is that pure translation technology providers have embarked on introducing AI copywriting tools, a strong endorsement of leveraging LLMs for multilingual content creation.
In 2023, the top TMS players were cautious as many AI-powered features released across different players were at first based on OpenAI, which is relatively easy to replicate and increases the risk of rapid obsolescence.
Source: LLM Solutions and how to use them, article and webinar by Nimdzi Insights
In 2024, it’s not only about OpenAI anymore. We have tracked the following common AI-powered features of TMS providers.
- LLM available as an MT engine
- LLM-based Quality Assurance
- Creative assistance for linguists (rewriting, alternative translations, SEO optimization, etc.)
- LLM-based Quality Estimation
- Content filtering (e.g. biased or harmful content)
- Fine-tuning of GPT
- Automated post-editing
- LLM-driven knowledge base
- Automated pre-editing
- LLM as a contextual engine for translation (not OpenAI)
- AI chatbot
- AI Content Generation
- LLM used as language resource (like TM)
While the above 13 AI-powered features are the most common in the TMS space as can be seen in Nimdzi’s TMS Feature Explorer, the list of implementations continues to grow. There are already other options, of course. To name a few,
- Bureau Works have also added automated term extraction and alignment in runtime, AI Tag Fixing, Smart Dynamic Feed — LLM-driven task priorities for users.
- GlobalLink has added Machine Interpreting to their product offering.
- memoQ has been developing Adaptive Generative Translation (AGT), an AI-based translation automation technology. It uses an LLM to generate translations and provides instant domain adaptation so that the translation results are tailored to the customer’s existing language resources (TMs, termbases, aligned documents).
- Phrase’s Automated Asset Curation (ability to clean language assets like TMs) and Custom AI capability to train custom MT engines are AI-powered.
- RWS already won an AI Breakthrough Award with their Evolve. Evolve combines a TMS (Trados Enterprise), RWS’ dedicated NMT (Language Weaver) which generates initial translations then subsequently evaluated by a trained Machine Translation Quality Estimation (MTQE) model, a private LLM, and language specialists. The quality estimation model here assesses and grades the quality of the output to determine which text segments need further editing. Next, content deemed adequate or poor by the MTQE system undergoes up to three iterations through a fine-tuned private LLM. The remaining segments are fixed by language specialists. Edits, either from the LLM or humans, are sent back to the NMT for adaptation and improved output.
- Smartling added fine-tuning of more LLMs in addition to GPT: Google Vertex AI, Bedrock, and IBM WatsonX hosted models.
- Smartcat announced their TMS platform is evolving to include AI dubbing and a subtitle editor. What is a standalone product in other, non-TMS, solutions is set to be a feature of Smartcat.
- XTM has added AI-driven TM matching optimization (Weighted Token Levenshtein, WTLV), monolingual and bilingual terminology extraction, TM alignment, and automatic placement of inline tags in the editor.
Last but not least, not only TMS providers are adding AI-driven features. For example, Pangeanic released ECO, a fine-tuned LLM primarily used for automated post-editing.
Translation Management System, reinvented
The TMS market is considered saturated by many, although newcomers help to shake up the incumbents. Nevertheless it is growing, although slower than other language technology segments. According to Jourik Ciesielski’s estimate, “the market continues to experience substantial growth, resulting in an estimated market size of USD 0.3 billion.”
For the start of 2024, the market size was revised to USD 321 million, according to research by Konstantin Dranch and Jourik Ciesielski, and the opportunity for these technology companies is bigger than was envisioned before. So how is the perception of the TMS market and its actors changing to facilitate this growth? Let’s recap.
It all started with the term CAT-tool (Computer-Assisted or Computer-Aided Translation), which was the main software to work on translations in the ‘90s and 2000s. CAT software already had such components as a bilingual editing environment, a TM, a termbase, and built-in quality assurance. But over time, to get the translation job done faster and make it more scalable, those components were no longer enough. That’s how a variety of business management features appeared in a CAT environment, resulting in the birth of TMS. For easier and quicker performance, many CAT tools emerged in the cloud and remain the essence of a TMS.
After 2010, this sector has been growing, with dozens of TMS being pushed to the market yearly. And the term CAT-tool has still been used, especially by linguists. But the enterprise sector, investor companies, and many LSPs are no longer talking about investing in CAT-tools. It's usually all about TMS — for the past five years at least.
"With the AI boom, TMS providers have been trying to reinvent themselves and change the perception of what this software can actually do for global companies. The focus has already shifted from just facilitating translation work to providing comprehensive content platforms. Some also went for the Language Operations (LangOps) concept to highlight the idea of an all-in-one solution for global content."
Yulia Akhulkova
TMS with AI capabilities are trying to be seen as strategic assets rather than just operational tools, as they aim to play a crucial role in global business strategies and competitive advantage. Their value propositions now emphasize the integration of AI to provide strategic insights for global businesses. A few examples:
- Smartcat is now “One AI translation platform for content in any language”. As they note, “Smartcat covers all your language needs with AI translation, AI content generation and AI human workflows”
- Crowdin leads with AI right in their motto “Three-Step Collaboration: AI, Human, Crowdin”
- RWS is a “provider of technology-enabled language, content and intellectual property solutions” which is “transforming content through translation, localization and AI-enabled technology – blended with human expertise”
- Unbabel is “the first Language Operations Platform fueled by an always-on AI that empowers you to bring in human review when needed”
- Lilt suggests to “generate content in any language with LILT Create” via “combining historical preferences with real-time user prompts”
- Smartling is a “LanguageAI platform to accurately translate and localize content into any language and any media at scale”
- Transifex suggests users “lift language barriers and engage the world with AI empowered translation and localization”
- Gridly was initially a CMS platform for game and software localization and grew into a comprehensive TMS. Their “Content and localization in sync” motto paints a vivid picture of all content pieces together, with translation reserved for AI bots:
Source: https://www.gridly.com/
This strategy pays off. With more AI integrations, TMS providers are increasingly attracting investment and interest from venture capitalists and enterprise sectors, highlighting their growing importance in the global market.
That’s how AI is transforming TMS from traditional translation and localization tools into sophisticated platforms that offer advanced capabilities and strategic benefits. The term “TMS” is being reinvented and we will most likely see even more new terms reflecting the added value proposition of the TMS providers in 2025 and beyond.
AI & Interpreting
When it comes to interpreting, human interpreters are not going to be substituted just yet, but AI will fill the gaps where no interpreting was offered before, allowing speakers to return to their native tongues and helping people get access to data in a fast and cost-effective way. However, without a universally accepted quality standard for human interpretation, it is difficult to determine how AI interpretation truly compares to a human interpreter. The key requirements for AI interpreting (or machine interpreting, MI) are not only quality but also latency, which is critical to the speeds needed for real-time speech translation.
The challenge lies in understanding when to express emotions and how to express them across cultures, and AI still struggles with that. However, there are some great improvements that AI can bring to the interpreting table, with its capacity to analyze and compare lots of data in a short time. "The tool is just as good as the data it is fed” stays true in this case. It is especially important for the possibility of perpetuating hateful or discriminating terms and language, and AI works great with safety measures, such as word-detecting systems or forbidden terms lists, outperforming its human counterparts in this area.
"While the Interpreting Systems category remains alluring with new entrants vying for prominence, the industry still awaits a defining breakthrough. Despite the buzz, the expected advancements have yet to materialize, and traditional metrics still dominate. This apparent inactivity hints that something significant is brewing beneath the surface."
Ewandro Magalhaes
So, while AI Interpreting is not going to outperform human translators now (or in the next couple of years), it is definitely going to be part of the landscape, serving human interpreters to provide even better results. The well-known solutions for MI include both devices (such as Timekettle’s and Waverly Lab’s products, or simple Google Pixel Buds) and software applications. The advantage of software applications is that they can be used within video conferencing interfaces, i.e., users can pull them into a Meet, Teams, or Zoom meeting. Nimdzi’s MI Evolution Matrix sees Kudo, Wordly, and Interprefy as leaders in this space.
Source: Evolution of Machine Interpreting, Nimdzi Insights
Quality evaluation versus quality estimation
While we talk about the quality of AI outputs for various applications, from ASR to localization, it is very important to differentiate quality evaluation from quality estimation. We dedicate different subcategories of our radar to these technologies, as they are quite different:
- Quality evaluation is a post-translation process. It is needed, for example, when reviewers want to determine how good the translation is or how long it would take to review it, which implies evaluating the linguistic quality.
- Quality estimation (QE) is the process that takes place at the time of translation. It is usually meant to predict the quality of MT output without human intervention. The automatic estimations/QE scores are used for instance to indicate whether raw MT can be used without post-editing.
On the QE front, there are two key players: TAUS and Modelfront. TAUS QE is a semantics-based quality score, telling users how close in meaning the two segments are. TAUS sentence embedding models are used to calculate this similarity score. For quality estimation, they offer generic and custom models. In the beginning of 2024, TAUS released an upgraded version of its quality estimation tool, featuring improved metrics, enabling a better correlation with human evaluations and a higher level of accuracy in translation quality estimation. Recently, TAUS published a new demo interface of the Estimate API. It's an interface where people can upload their documents and have their content scored with TAUS QE scores. It is currently available in English to Spanish, French, German, and Italian.
Source: https://www.taus.net/
Another key actor in the quality estimation market is Modelfront. What they mainly do is predict which machine-translated segments do not need post-editing. This helps significantly increase human linguist throughput. To predict whether an MT segment is good or bad, ModelFront API learns from post-editing data, to reflect domain, terminology, and style. They support any combination of more than 100 languages out of the box.
"MT quality estimation has evolved from a nice-to-have feature in 2023 to a language technology category in 2024. Accelerated by the dim economic situation, it enables various localization stakeholders to make quick, technology-driven decisions about which content to prioritize for translation, or which workflow to apply to both high-quality and low-quality machine translations. This is an excellent opportunity for localization managers with lots of MT-ready content — think about e-commerce product owners."
Jourik Ciesielski
Where does language technology come from?
Major technology companies are headquartered in the US. They, however, set an example for other countries. For instance, consumer-level success helped Google create a competitive advantage in MT that was transferred to their API-based MT solutions. This example was followed to success by European companies such as DeepL (Germany) or Translated (Italy), who both launched easy-to-use, high-quality automated translation solutions that were picked up by consumers and small businesses, and transferred their visibility to serve enterprise customers.
OpenAI followed a similar path with their LLMs, launching ChatGPT to the wider public that was transferred to business users with GPT-4 via their own and Microsoft’s application layers and infrastructure. The emergence of European players such as Mistral and Silo can potentially help achieve similar success in the LLM arena as that of DeepL.
Nevertheless, US tech giants dominate the language technology market and development space in all major categories of the Nimdzi Radar. Big tech companies are in the market despite the fact that language is not their main business.
Map of language technology; data from over 660 technologies
The result is a centralized language technology market dominated by Big Tech servicing large enterprise demand, with a long tail of language, locale, industry vertical, and function-specific language solution providers that essentially rely on Big Tech’s ability to drive the expensive core innovations. The general trend is that a handful of foundational models are used by the majority of companies on our radar to compile their products and applications, as the development of the foundation models is complex, complicated, and expensive.
The training of the most popular LLMs is skewed towards English, especially because English is the lingua franca of technologies in general. Many benchmarks are created in English, English data is the most available for training, and many of these tools are developed by US companies with American customers in mind.
While content in other languages exists within these datasets, the percentage and quality vary. This poses several risks. Beyond the potential loss of cultural language heritage, it also creates a situation where LLMs perform poorly for less-resourced languages (such as Estonian or Maltese in the EU, as well as many African and South-East Asian languages), reinforcing existing digital divides. Many language tools work well only for certain language pairs out of the possible combinations (e.g., an MT solution may work well between English and Maltese, but not from Maltese to Irish and Estonian to Portuguese, which can be critical for Europe with its diverse set of languages).
The performance of solutions for monolingual language tasks (speech recognition, text generation, or summarization) varies greatly between languages, as well, — especially with the major commercial LLMs. One of the primary factors for this variation is, again, the availability of data in a specific language. In response to these concerns, various initiatives are emerging to gather data in non-dominant languages.
The future outlook
Machine translation post-editing
The continued evolution of neural network architectures, particularly transformer models, will lead to more accurate and contextually aware translations. These models will likely reduce the need for extensive human intervention, as future MT systems will increasingly incorporate sophisticated contextual understanding and domain-specific knowledge, leading to translations that are not only grammatically correct but also contextually appropriate. This will reduce the frequency and extent of errors that require post-editing.
Some MT systems already incorporate real-time learning capabilities where they adapt based on user feedback and correction patterns. This adaptive learning will continuously improve translation quality and make it more reliable, thereby decreasing the need for post-editing. Some companies on our radar offer automatic post-editing, others (e.g. Bureau Works) are already using LLMs to answer the question of “considering the MT, the TM, the glossaries, and the changes the translator has already made, what is the best suggestion to make for this segment?”
Moreover, improved preprocessing techniques, such as better handling of idiomatic expressions and cultural references, will enhance translation accuracy before it reaches the post-editing stage. This proactive approach will reduce the need for extensive post-editing.
"We predict that more advanced quality assurance algorithms will emerge, capable of automatically detecting and correcting a wide range of issues that currently require human intervention. These solutions could handle many of the tasks traditionally associated with Machine translation post-editing (MTPE), such as error detection and correction. As MT and LLMs quality improves, linguists may shift their roles from post-editing to more strategic tasks like high-level oversight, content verification, and sanity checks. Their expertise will be used to ensure that the content in the target language aligns with the intended meaning, tone, ethical and cultural nuances, rather than correcting linguistic errors."
Josef Kubovsky
While MTPE as an end-of-pipeline quality step will not disappear entirely, its necessity is likely to diminish as MT technology continues to advance, and human oversight may turn more towards verification and validation of pre-translated high-risk content.
“AI, AI, AI”
The rise of generative AI is indeed poised to transform the language technology industry in significant ways. Businesses everywhere are under increasing pressure to demonstrate how they are harnessing the power of AI to drive efficiencies and scale production. The spotlight is on generative AI and LLMs to deliver right away.
"Major consultancy firms are pushing forward inflated estimates of the corporate profits that could arise from implementing generative AI with current state technology. These overly optimistic projections seem to be more effective at attracting investors than at convincing enterprises to adopt the technology. In reality, the early adopters have already embraced the tech; the late majority will take their time, especially due to concerns over data privacy and security (many without clear data governance frameworks), difficulties with adoption and resistance (unclear adoption plans that lack a holistic approach), and most importantly (especially now), fears surrounding regulatory compliance."
Nadezda Jakubkova
At the same time, “AI” appears to be not only the most frequently used term in our report but also a catalyst for language technology development at least in two significant ways:
- The hype around LLM solutions triggered a new wave of interest and investment into language technology. Startups and incumbents aim at augmenting and replacing existing language tools and workflows, as well as at creating new solutions to problems that were previously not possible or feasible to solve.
- LLMs brought language technologies into the spotlight for the executive level. New ways of productivity gains are always prized highly in competition, and language AI suddenly became both widely available and easy to start with.
"Now that the hype around GPT translation features and AI assistants in TMS has cooled down, and prompt engineering did not prove to be as big a differentiator as anticipated, LLM fine-tuning has taken center stage. Many tech-enabled LSPs will smell an opportunity given the large open-source community driven by Meta, the relative ease of fine-tuning a model, and the numerous papers reporting that fine-tuned models outperform Amazon, Google, Microsoft, and closed GPT-4."
Jourik Ciesielski
So where do we go from here?
Before we close our annual analysis, let’s use a comprehensive graphic that visualizes how we see the language industry has been progressing in the past and what we see in the next 10 years.
Source: The Language Industry Curve by Renato Beninatto and Laszlo K. Varga
This graph prepared by Nimdzi’s own Renato Beninatto and Laszlo K. Varga allows us to grasp the layered significance of technological advancements, infrastructure developments, and key industry events that have shaped and continue to shape the language industry. It provides a comprehensive overview of the past, present, and projected future of this dynamic and rapidly evolving sector.
"There is a lot of hype at the beginning of every innovation, but in the real world, the adoption of new technology is far more nuanced and gradual than people often predict. Just because a technology is accessible and simple to use doesn't guarantee immediate and widespread adoption. The reality is shaped by industry-specific needs, generational influences, and the complex interplay of legal and regulatory frameworks. Understanding these factors is crucial for driving innovation effectively, especially in fields like localization where the integration of AI, including GenAI, must be approached with careful consideration and expertise."
Renato Beninatto
We anticipate that while new advancements in technology, communication, and infrastructure will continue to challenge the status quo, the language industry will continue to grow by adapting to changing circumstances with the resilience and flexibility it has demonstrated with every major external event in the past.
Language technology solutions and products will continue to evolve. We can also expect that once the AI hype cools down, the “AI” prefix in company names and value propositions, now considered obligatory, will be dropped — just as it happened before with many other developments (remember “big data” or “cloud”?).
And to continue tracking and reflecting the changes in our industry, we invite all language technology developers to join hundreds of companies in the growing curated catalog of language technologies on the Nimdzi Language Technology Radar.
The report was published by Yulia Akhulkova, Nimdzi's Technology Researcher, on September 1, 2024.
If you have any questions about language technology, reach out to Yulia at [email protected].
