OnlineDocTranslatorGuides

Answer

Machine Translation Software for Certified Translation Production

Machine translation software produces first-draft output that human translators post-edit inside CAT grids before certified delivery. Engine choice, LQA gates

Machine translation software sits inside certified translation production as a first-draft stage, not as a signing authority. The engine runs after translation memory and before human post-editing, inside a CAT tool, and feeds a fixed pipeline that ends at an ISO 17100-aligned LQA gate. The parent reference for the human stage that follows is Post-Editing Machine Translation: Rates, Standards, and Workflow.

Key points

  • Machine translation software produces a first-draft only; certification requires a human post-edit inside a CAT grid against a locked term base.
  • The fixed certified-lane pipeline order is translation memory → machine translation → human post-editing (MTPE) → AI proofread → LQA gate — proofreading never runs before post-editing.
  • Full post-editing bills at 50–70% of the full human word rate; light post-editing at 30–40%; MTPE against an untested engine bills at the full human rate (100%).
  • The LQA gate runs at $40–$90 per 1,000 words against an MQM-style scorecard aligned to ISO 17100 — below that floor, MT savings stop compounding.
  • Public free MT endpoints are disqualified for USCIS, court, and medical lanes; only on-prem engines, private APIs with no-training contracts, or DPA-covered accounts clear regulated work.

How it works

What is machine translation software in a certified translation workflow?

1

How does machine translation software differ from a CAT tool?

Machine translation software generates automatic output from a source segment; a CAT tool is the working environment that holds translation memory, termbase, QA checks, and the segment editor — and plugs an MT engine in as one suggestion source among several. Google Translate is an MT engine, not a CAT tool. The CAT tool stays the working surface; the engine provides draft suggestions the translator accepts, edits, or rejects. The canonical reference on this split is CAT Tools for Translation: What They Do and Which Ones Translators Actually Use.

2

Where does MT sit in the TM → MT → MTPE → LQA pipeline?

Machine translation runs second — after translation memory fills exact and fuzzy matches, before human post-editing, and well before the AI proofread and the LQA gate. Pipeline order is fixed on certified lanes: proofreading raw MT polishes wrong sentences into fluent wrong sentences. The sequence is documented in full on Translation Processes: How Agencies Run Certified Production.

flowchart TD
 A[Source file] --> B[Translation memory fills matches]
 B --> C[MT engine drafts remainder]
 C --> D[Human post-edit in CAT grid]
 D --> E[AI proofread]
 E --> F{LQA gate $40-$90 per 1k words}
 F -- Pass --> G[Certified signer attestation]
 F -- Fail --> D

Which machine translation engines do translators actually plug into certified production?

Production translators plug DeepL, Google Translate, Microsoft Translator, Amazon Translate, SYSTRAN, and LLM-based engines (GPT, Claude, Gemini) into their CAT tool or TMS. The engine is chosen per language pair and per lane by BLEU/COMET scores, domain fit, and measured post-edit distance on recent jobs — never by vendor marketing.

How do DeepL, Google Translate, Microsoft Translator, Amazon Translate, and SYSTRAN compare for professional use?

DeepL leads on tone and nuance in its core European pairs; Google Translate leads on language coverage; Microsoft Translator suits Microsoft-centric stacks; Amazon Translate fits high-volume AWS builds; SYSTRAN handles technical content with domain-trained engines. None is universally best. The choice is per language pair.

EngineStrengthTypical lane fit
DeepLTone and nuance in core European pairsMarketing, legal in EN/DE/FR/ES/IT/NL
Google TranslateBroadest language coverage; CAT plug-inLow-resource pairs, comprehension drafts
Microsoft TranslatorOffice formats, Microsoft-centric stacksEnterprise teams on Microsoft 365
Amazon TranslatePer-character pricing at AWS volumeHigh-volume software strings on AWS
SYSTRANDomain-trained engines for technical contentPatent, pharma, engineering
LLMs (GPT, Claude, Gemini)Tone control and glossary-following via promptSuggestion source, not a certified signer

How do large language models (GPT, Claude, Gemini) fit as translation engines?

Large language models function as general-purpose translation engines with strong tone control and glossary-following via prompt, but with inconsistent output across runs and weaker guarantees on confidential data than dedicated MT APIs. On certified lanes they are a suggestion source, not a signing authority.

Which engine should a translator pick for a given language pair?

Pick the engine with the lowest measured post-edit distance on a recent job in that pair and domain, not the engine with the highest advertised language count. The selection rubric runs in this order:

  1. Pull three recent files in the target pair and domain.
  2. Run each candidate engine against them inside the CAT grid.
  3. Measure post-edit distance (edit count ÷ MT word count).
  4. Score BLEU and COMET on the same sample.
  5. Check confidentiality posture for the lane (public endpoint or private API).
  6. Pick the engine with the lowest post-edit effort that clears the confidentiality gate.

For multi-lane programs where the engine choice differs per language, see Localization Solution: What Translators and Agencies Assemble to Run Multi-Lane Work.

Pricing

What does machine translation software cost inside a translator's rate card?

The engine API itself is a per-character or per-word cost of cents, but the real cost is the human post-edit and LQA gate on top. Full post-edit bills at 50–70% of the full human word rate; light post-edit at 30–40%; MTPE on an untested engine bills at the full human rate; LQA runs at $40–$90 per 1,000 words.

Line itemPricePrecondition
Light post-edit30–40% of full human word rateEngine tested on client's own strings
Full post-edit50–70% of full human word rateEngine tested on client's own strings
MTPE on untested engine100% (full human word rate)No domain test cleared on client strings
LQA gate$40–$90 per 1,000 wordsMQM-style scorecard aligned to ISO 17100
USCIS certified translation$0.10–$0.25 per source word or $25–$150 per 250-word pageQualified signer and certificate of accuracy

How is MTPE priced as a line item on a certified quote?

Post-editing is quoted as a split line: a raw-MT line plus a human post-edit line. The combined price runs 50–70% of the full human word rate for full post-edit and 30–40% for light, provided the engine was tested on the client's own strings first. For what the post-editor delivers at each tier, see Post-Editing Translation: Workflow Standards and What Editors Are Expected to Deliver. Agencies positioning MTPE as a service line should review Machine Translation Post-Editing Services: What Agencies Offer and How to Position Yours.

Why does MTPE on an untested engine bill at the full human rate?

When the engine has not been tested on the client's own strings, the post-editor absorbs the quality risk of unknown engine behavior. Billing the discount rate on that risk is how post-editors lose money. The full human rate (100%) is the correct price until a domain test clears the engine for the client and the lane.

American Translators Association Certification: Exam, Eligibility, and Credential Recognition

What quality controls does machine translation software need to pass before certified delivery?

Every MT-drafted file passes four controls before delivery: a human post-edit against the locked term base, an AI proofread that runs only after the human pass, an LQA gate at $40–$90 per 1,000 words scored on an MQM-style scorecard aligned to ISO 17100, and a certified signer attestation.

  • Bilingual human post-edit inside the CAT grid against the locked termbase.
  • AI proofread after the human pass, never before — polishing raw MT ahead of the human stage hides fluent errors.
  • LQA sample scored on an MQM-style scorecard aligned to ISO 17100.
  • Certified signer attestation on the translator's certificate of accuracy.

What does the LQA gate check, and what does it cost?

The LQA gate samples translated output at $40–$90 per 1,000 words against an MQM-style scorecard aligned to ISO 17100, scoring accuracy, fluency, terminology, locale convention, and style. Files below threshold return to the post-editor. LQA is non-negotiable on certified lanes. Below the LQA floor, MT "savings" stop compounding — the real break-even for adopting a new engine is whether its post-edit output clears LQA on the first pass.

How does ISO 18587 constrain machine translation post-editing?

ISO 18587 sets the competence requirements for post-editors and the quality bar for full post-editing of MT output. It is the standard that lets an agency bill MTPE as a certified deliverable rather than an unreviewed MT dump. ISO 17100 sets the surrounding translation service quality requirements — the standard the LQA scorecard aligns to.

About

How do translators handle confidentiality and data residency when using machine translation software?

Public MT endpoints send source text to the vendor and may train on it, which disqualifies them for USCIS filings, court documents, and medical records. Certified lanes require a private API with a no-training contract, an on-prem deployment, or a vendor with an explicit confidentiality clause and encryption in transit and at rest.

Which MT deployment modes are safe for USCIS, court, and medical lanes?

Safe modes include on-prem engines, private cloud deployments with no-training contracts, and vendor accounts under a signed DPA with encryption at rest. Free public web endpoints are not safe for regulated lanes and are refused on this site's certified pipeline.

Deployment modeUSCISCourtMedical (HIPAA)
Free public web endpoint (Google Translate web, ChatGPT free)DisqualifiedDisqualifiedDisqualified
Paid API tier, standard termsCase-by-caseDisqualifiedDisqualified
Paid API with no-training contract + DPAClearedCleared with DPACleared with BAA
Private cloud deploymentClearedClearedCleared with BAA
On-prem engineClearedClearedCleared

Pricing

How does machine translation software integrate with CAT tools, TMS platforms, and glossaries?

Machine translation software integrates as a plug-in inside CAT tools (Trados, memoQ, OmegaT, Phrase, Smartcat) and as a configurable layer inside TMS platforms (Crowdin, Lokalise, Smartling, Transifex). The CAT tool remains the working environment. The MT engine provides suggestions that the translator accepts, edits, or rejects segment by segment, with the glossary enforced before submission.

American Translators Association Certification: Exam, Eligibility, and Credential Recognition

When does machine translation software not belong in a certified lane?

Machine translation software does not belong on handwritten source files, on low-resource language pairs where post-edit distance exceeds a fresh human translation, on dialect-sensitive legal text, or on any lane where the engine has not been domain-tested on the client's own strings. On those files, the engine is turned off and the translator works directly from source. The rule is straightforward: when the law, a life, or a client's financial exposure is at stake, relying on an untested engine creates risk that no post-editing workflow can reliably neutralize.

MT output is a first-draft tool, not a signing authority. Its value inside a certified production pipeline is strictly proportional to how well the selected engine has been validated against the specific domain, register, and language pair in question. When that validation has not been performed — or when the content category inherently resists it — the engine adds noise rather than speed, and the production lane should default to full human translation from the outset.

  • Handwritten civil records. Birth, marriage, divorce, and academic records in handwritten form produce fragmented OCR output that an MT engine cannot parse; errors compound through every subsequent step.
  • Low-resource language pairs. When measured post-edit distance on a sample of the client's own strings exceeds the effort of a from-scratch human translation, the engine delivers negative productivity and must be excluded.
  • Dialect-sensitive legal text. Jurisdictional register is not a stylistic preference — it is a legal requirement. Text requiring a specific dialectal or jurisdictional register must be handled by a qualified human translator without MT intervention.
  • Untested client files. Any file type or domain the engine has not been benchmarked against using the client's actual content is an untested engine for that client, and must be priced and staffed at the full human word rate.
  • High-stakes content with no tolerance for error. Content where a single mistranslation carries legal, medical, or material financial consequences requires human translation as the primary production method, with MT excluded from the lane entirely.

Applying these exclusions is not a limitation on MT adoption — it is the governance layer that makes MTPE credible in the lanes where it does belong. A disciplined exclusion policy protects the certification, the client, and the production team simultaneously.

How we run machine translation software on this site

We produce the machine translation first draft, and then a qualified human translator post-edits it inside the CAT grid against the locked term base. The AI proofreader runs after the post-edit, never before — polishing raw MT ahead of the human stage is how fluent errors reach the LQA gate.

We decline to price MTPE against an engine we have not tested on the client's own strings. We have watched post-editors absorb translation risk at a post-editing rate too many times. If the engine has not cleared a domain test on recent client files, we bill at the full human word rate until it does.

We decline monolingual post-editing on certified lanes. Certified lanes require a bilingual post-edit against the source — reading the MT output alone cannot catch mistranslation, omission, or false cognates. We also refuse free public MT endpoints on USCIS, court, and medical files, regardless of how fluent the output reads.

Working translators ask four questions repeatedly about machine translation software: which engine is best, whether free tools are usable professionally, whether LLMs replace NMT, and whether MT output can be certified. Each has a short answer below.

Frequently asked questions

Which AI tool is good for translation?

For certified translation production, the best AI tools are purpose-built machine translation engines integrated into CAT workflows — such as DeepL API, ModernMT, or dedicated certified-translation SaaS platforms. General-purpose tools like ChatGPT can produce fluent text but lack the consistency, terminology control, and audit trails that certified translation requires. Professional translators and agencies benefit most from tools that support translation memories, glossaries, and human post-editing pipelines, ensuring the final certified output meets legal and regulatory standards.

Is ChatGPT a machine translation?

ChatGPT is a large language model (LLM) that can translate text, but it is not purpose-built machine translation software. Unlike dedicated MT engines (DeepL, Google Translate API, ModernMT), ChatGPT has no built-in translation memory, glossary enforcement, or segment-level quality scoring. For certified translation production, this matters: professional workflows require consistency across repeated terms, version control, and a defensible post-editing process — features that specialist MT platforms provide and that general LLMs lack out of the box.

Is DeepL a good translation tool?

DeepL is widely regarded as one of the highest-quality neural MT engines available, particularly strong for European language pairs. Its fluency and grammatical accuracy make it a popular choice for post-editing workflows, reducing the time professional translators spend on first-draft revision. However, for certified translation production, DeepL alone is insufficient — translators still need CAT tool integration, terminology management, and a formal certification step. DeepL Pro's API enables these integrations, making it a solid MT backbone within a broader professional workflow.

How much does DeepL cost per month?

DeepL Pro plans start at around $10–$12 per month for the Starter tier (individual use, limited text volume), rising to roughly $25–$30/month for the Advanced plan and $60+/month for Ultimate, which includes higher API limits and CAT tool plugins. API access for developers and agencies is usage-based, priced per character translated — around $25 per million characters. For agencies processing high volumes of certified translation drafts, usage-based API pricing becomes more economical than flat-rate plans.

What is the best machine translation tool?

The best machine translation tool for certified translation production depends on language pair, volume, and integration needs — but leading engines include DeepL (strong on European languages), Google Cloud Translation API (broadest language coverage, 100+ languages), and ModernMT (adaptive MT that learns from translator corrections in real time). For professional and certified contexts, the "best" engine is one that integrates cleanly with CAT tools, supports custom glossaries, and feeds into a structured post-editing and certification workflow rather than being used in isolation.

What are computer-assisted translation tools?

Computer-assisted translation (CAT) tools are software applications that help professional translators work faster and more consistently by breaking documents into segments and storing previously translated text in a translation memory (TM). Popular CAT tools include SDL Trados Studio, memoQ, Wordfast, and Memsource (Phrase). They differ from raw machine translation in that the human translator remains in control — the CAT tool surfaces MT suggestions and TM matches, while the translator post-edits and approves each segment. For certified translation production, CAT tools also help maintain terminology consistency across legally sensitive documents.

Is a translator a well paid job?

Translation pay varies widely by specialization, language pair, and work type — but certified translation is among the higher-paying segments. Freelance certified translators in the US charge $0.12–$0.25 per source word, with legal and technical documents commanding premium rates. Full-time in-house translators at agencies earn roughly $45,000–$75,000 per year depending on location and specialization. Translators who leverage machine translation with post-editing can process higher word volumes per day, improving effective hourly earnings while maintaining quality standards required for certification.

What are some popular software tools used for translation?

The most widely used professional translation tools fall into two categories: CAT tools and MT engines. Leading CAT tools include SDL Trados Studio, memoQ, Phrase (Memsource), and Wordfast. For machine translation, DeepL API, Google Cloud Translation, and ModernMT are common choices. Certified translation workflows often combine both — an MT engine generates a first draft inside a CAT tool, a professional translator post-edits it, and the agency certifies the final output. Dedicated SaaS platforms for certified translation production integrate these layers into a single, audit-ready pipeline.

What are some AI translation tools?

Leading AI translation tools used by professional translators and agencies include DeepL (neural MT with high fluency), Google Cloud Translation API (widest language coverage), ModernMT (adaptive engine that improves with translator feedback), Amazon Translate, and Microsoft Azure Translator. For certified translation production specifically, the key differentiator is whether an AI tool can be embedded into a post-editing workflow with terminology control and audit trails — not just whether it produces fluent output, since certified documents carry legal weight and require human verification.

What is the best translation engine?

No single translation engine leads across all language pairs and document types — the best choice depends on your specific use case. DeepL consistently scores highest on fluency evaluations for European languages; Google Cloud Translation API offers the broadest coverage (100+ languages); ModernMT adapts to a translator's own style and terminology over time, making it particularly valuable for certified translation production where consistency is critical. Agencies producing high volumes of certified translations benchmark multiple engines on their actual document types before committing to one.

Is machine translation possible?

Yes, machine translation is not only possible but widely deployed — modern neural MT engines produce drafts that are fluent enough for professional post-editing in most language pairs. For certified translation production, MT is used to generate a first draft that a qualified human translator then reviews, corrects, and certifies. This human-in-the-loop model — often called MTPE (machine translation post-editing) — is now a standard industry practice, recognized by ISO 18587, and allows translators to process significantly more content per day without sacrificing the accuracy that certified documents require.

Related