Generic MT vs Custom MT Engines: How to Choose the Right Fit for Your Content

Machine translation has become an essential part of modern localization workflows. It speeds up delivery, reduces repetitive effort, and supports large-scale content production. But there is still a recurring question for LSPs, vendors, and end clients alike:

Should you rely on a generic MT engine or invest in a custom-built one?

There is no universal answer. Both options can deliver strong results in the right conditions and disappointing results in the wrong ones. The key is understanding how the engines differ, what each one is designed to handle, and how your own project variables influence the final output.

This post takes you through the main considerations, using insights from TextTrans’ specialists and examples from our Adventures in Machine Translation video series.

Understanding the difference between generic and custom MT

Some MT engines are built to handle almost anything. Others are built to handle one thing well.

A generic engine relies on large multilingual corpora and no domain tailoring. It tends to perform reasonably well with simple instructions, consumer-facing text, or general information.

A custom engine is trained with in-domain data. It can learn tone, conventions, and terminology specific to an industry, sometimes to a level that brings MT output closer to a human draft.

Neither option is universally better. Each has strengths, limitations, and real-world costs.

Before choosing: what really matters

The most reliable way to evaluate which MT engine fits your project is to look at four core factors:

Content type

Some content types work well with generic MT. Others do not. Highly technical, legal, financial, or brand-driven text benefits from domain training. More straightforward text may not require it.

Language pairs

Not all language pairs perform equally. English→Portuguese, for example, can introduce challenges with gender agreement, syntax, and segment linkage. Domain training helps stabilize these patterns.

Time available

Custom MT requires development cycles, testing, and refinement. If you need an immediate solution, a generic engine combined with skilled post-editing may be more practical.

Budget

Custom engines require investment in training data, evaluation, and maintenance. For high-volume content in a consistent domain, the return is strong. For small or mixed content, generic MT may be enough.

Each factor shapes the engine choice. The more specialized your content and constraints, the more carefully this decision needs to be made.

Why feedback from linguists matters

Translators and post-editors are the first to notice what an MT engine consistently gets wrong and what it handles well. Their feedback reveals:

• recurring error patterns
• terminology issues
• grammar habits of the engine
• tone and register weaknesses
• time spent fixing specific types of content

This insight is essential for deciding whether to keep using the same engine, switch, or move toward a custom model. No metric offers clearer real-world evidence than the feedback of the people working with the output.

Common MT errors: what the team sees in practice

Episodes 5 and 6 of our series focus on the types of mistakes MT engines tend to make with more complex content. Some are mechanical. Others distort meaning or disrupt the flow of the text.

Words with multiple meanings

Patricia Freitag notes that everyday words can cause disproportionate issues. A word like chip can refer to a snack, a hardware component, a digital token, or part of a payment card. Engines that lack domain cues often guess incorrectly.

Punctuation inconsistencies

Spacing around abbreviations like .pdf often leads to errors. These are small issues but appear across many files, adding unnecessary correction time.

Product and brand names

Some engines mistranslate or modify product names that should remain untouched. Well-known brands such as YouTube or AirPods may appear in altered forms.

Incorrect segment linkage

This is especially common in English→Portuguese. MT may treat each segment independently without understanding the logical sequence, producing text that reads correctly line by line but is incoherent when read as a whole.

Errors like these illustrate why MT cannot replace human judgment and why post-editing remains essential.

When custom MT makes a difference

A custom MT engine earns its value in situations where precision matters. This often happens in industries with strict terminology or fixed writing conventions. Financial institutions, legal teams, and highly technical manufacturers all fall into this category. When the content repeats similar concepts across long stretches of documentation, a custom engine absorbs that logic and begins producing text that aligns naturally with the domain.

It also helps when the brand has a recognisable voice. Marketing teams spend years shaping tone, rhythm, and messaging, and a custom MT engine can learn those patterns far better than a generic model. The improvement is rarely dramatic overnight, but through training cycles and sample reviews, the engine becomes more predictable. That predictability is what reduces post-editing time.

The real advantage emerges over long-term, high-volume workflows. Once the model is trained and stable, it becomes an asset that supports consistency across thousands of future files. It is not a universal solution, though. If the domain shifts or the content becomes more varied, the strength of the custom model begins to weaken. Custom MT only works when the content feeding it remains aligned with the content you expect from it.

When a generic MT engine is enough

There are many cases where generic MT produces usable output without needing heavy customization. Short, functional text often falls into this category. Instructions, labels, status messages, or UI strings that carry no brand weight and little narrative nuance tend to translate well through generic engines.

Generic MT also becomes the practical choice when projects move quickly and budgets are tight. In these conditions, teams need something that works immediately without lengthy testing. For small volumes, or for mixed content covering several unrelated topics, a generic engine may not only be sufficient but more efficient. It avoids the risk of training a custom model on data that will not be reused in the future.

The key is being honest about expectations. Generic MT is not meant to match the style of a seasoned marketing writer or replicate the precision of a legal translator. It offers speed and convenience, and when used for the right type of text, that is exactly enough.

The role of post-editing

Regardless of which engine you choose, post-editing remains an essential step. Machines can assemble sentences, but they cannot yet anticipate the subtle differences between intention and interpretation. This is where human judgment becomes irreplaceable.

Post-editors are the safety net for tone, accuracy, and flow. They correct misleading phrases, adjust grammar, reshape awkward structures, and confirm that terminology has been applied consistently. In brand-heavy content, they refine the voice. In technical content, they check the logic. When MT output struggles with segment linkage, the post-editor is the one who rebuilds coherence.

The amount of post-editing varies depending on the engine and the quality of the input, but the value of this step does not change. It is the moment when MT becomes usable.

How translators evaluate MT in real projects

Translators do not assess MT output in the abstract. They experience it line by line, under deadline pressure, inside real client workflows. Over time, they begin to recognise the “behaviour” of an engine.

Some engines tend to drift stylistically halfway through a paragraph. Others cling to literal structures even when the meaning becomes distorted. In Portuguese, for example, verbs occasionally migrate to unnatural positions because the engine has not absorbed how emphasis and rhythm work in the language. These quirks accumulate and shape how linguists feel about the engine as a whole.

A translator’s assessment is rarely a single judgement. It builds gradually, through patterns that reappear across different files, different domains, and different levels of complexity. Their insight is what reveals whether an engine is predictable, erratic, or simply not suited to certain content types.

A small illustrative snapshot:

Recurring signal	What it suggests
Sudden register shifts	Engine hasn’t learned tone or domain conventions
Inconsistent sentence flow	Weakness in handling segment relationships
Correct terminology but clumsy phrasing	Model trained on glossaries but not style
Stable grammar but wrong meaning	Lack of contextual anchors

These patterns tell a far richer story than raw BLEU scores ever could.

Making the final decision

Some teams try to decide between generic and custom MT as if they were comparing two software products. In practice, it’s closer to choosing a workflow philosophy.

If your content demands exact terminology and you work with the same client repeatedly, a custom engine becomes a long-term companion. It grows with the project. It learns. It stabilises. And over time, it reduces friction.

If your projects shift unpredictably between domains, a custom engine may never find its footing. In those cases, a generic engine becomes the more flexible choice, and the skill of the post-editor becomes the defining factor in quality.

For highly narrative, brand-driven, or high-stakes content, neither engine—generic or custom—can be expected to carry the weight alone. Human reasoning remains central. MT becomes a support tool, not a foundation.

Closing reflections

Machine translation works best when treated as a strategic choice rather than a shortcut. A generic engine can be fast and convenient; a custom engine can be accurate and dependable. Neither exists in isolation from the people who shape the final text.

The linguists who work with MT output every day are the ones who understand its strengths, its blind spots, and its limits. Their feedback should guide the selection, not metrics or assumptions.

In the end, MT succeeds not because it replaces human expertise, but because it enables that expertise to be applied where it makes the most impact. And that is where the real efficiency lies.

---