Custom.MT Console has rolled out a key update: a way to combine the output of popular machine translation engines with generative AI. Localization specialists may now easily style MT output and “mix and blend colors” to get a more accurate and evocative machine translation that speaks with the audience.

Machine translation engines are reliable but rigid, and generative AI rewrite gives them flexibility. Localization teams can now control the tone of voice, modality, terminology, punctuation, and translation length in translation across any machine translation engine. 

Use cases:

1.  Fix machine translation errors from any engine
2. Formal/informal translations, “toi” in marketing, “vous” in legal
3. Tone of voice
4. Custom prompts
5. Language variants: rewrite for UK/US English, rewrite for Latam Spanish
6. Rewrite for specific videogame character gender
7. Convert feet to meters, pounds to kg

Test it with immediate results: post-editing via Custom.MT console is available in Trados, memoQ, and Smartling, and it works across 90+ languages supported by OpenAI.

Here is what it looks like:

1. New Templates

The new Templates interface helps build workflows and sequences of MT & AI models. In the sequence, it’s possible to add GenAI both before the translation step to spell check and rewrite incoming source content for pre-editing and after it for post-editing of machine translation output coming from an engine.

Automated Pre-editing: simplify & improve readability, remove personal names, change units of measure

Automated Post-editing: change formality, tone of voice, correct translation errors.

The Template executes machine translation and prompts on the sentences it receives from the CAT tools segment by segment, however, it can also work on paragraphs and whole documents with context when called via an API.

Each use case requires a prompt that may be written in the Prompt Studio, which is another page freshly added to the Console.

2. Prompt Studio

In the Prompt Studiom, users may write and edit prompts for OpenAI, Gemini, and other large language models. It comes with 6 pre-built prompts for the original release, and we encourage users to create new prompts as soon as possible. 

Basic controls:

• Temperature setting – we recommend 0 or close to zero for repeatable output
• Model selector – we recommend prioritizing GPT-4 Turbo 1106 for the quality/cost/speed ratio

For advanced users, the Console includes the ability to add extra variables and linking them to an external source such as a Google Sheet/ Applications include a table of context such as character genders & character profiles for video game translations, links to screenshots for visual context, and perhaps even external databases for factual information retrieval.

Limitations

GenAI’s addition to the Console is a powerful new feature that provides nuanced control over machine translation output, and an easy way to fix recurring errors. It’s an improved and quickly learnable toolkit for project managers and localization engineers to make automated translation more accurate and flexible.

However, it is not a panacea. Here are some of the current limitations of the system:

1. GenAI sees the text segment by segment
   When Custom.MT console works with TMS and CAT tools, text for translation arrives in segments, and the models have to translate them one-by-one, without seeing the whole document for context. In the future, we hope to address this by providing previous segments as context in prompts via the variables system.
   
2. Users need some prompt writing skills
   Prompt writing is a skill with a learning curve. To achieve clean results without notes or explanations, use short prompts with examples of the output, and low temperature settings, otherwise, you might see certain models explaining their actions instead of just generating the required output. Be prepared to experiment and learn. Experience is a major factor in achieving better results with prompting. You can learn prompts that are practical in localization from specialized workshops such as this one.
   
3. This is the first step
   We’re rolling out the combo of MT and GenAI with a toolkit and a basic set of features that early adopters may immediately take advantage of. There is much more to build to make automated post-editing an integral part of every localization workflow. 

Availability & Costs

At launch, we’re making this functionality available on any business plan for two weeks. After the trial period, it will remain available to organizations on medium and higher-tier plans.

According to our calculations and the average length of prompts our users have written so far, adding GenAI steps to machine translation increases the costs by €10-40 per million characters with GPT-4  Turbo 1106 and by €50-100 when using plain GPT-4. These are approximate values based on prompts under 150 words.

Example calculation 1:

• Microsoft Translator – €10
• Custom.MT Console – €20
• OpenAI GPT-4 Turbo + medium prompt – €25

Total: €55 per million characters

Example calculation 2:

• Google AutoML Machine Translation – €75
• Custom.MT Console – €20
• OpenAI GPT-4 Turbo + long prompt – €40

Total: €135 per million characters

In most cases, using short prompts will be more affordable than fine-tuning the machine translation models and using the customized MT. For users who are looking for the best quality possible, we recommend doing both: fine-tuning to leverage years of accumulated translation memory, then scanning for recurring errors and fixing them with short prompts.

The function is immediately available in Trados, memoQ, and Smartling.