Machine Translation Post-Editing (MTPE) — using machine translation output as a first draft that a human editor then reviews and corrects — has become a significant topic in game localization. Advances in MT quality (particularly neural MT engines like DeepL and GPT-based systems) have made MTPE a more viable option for some content types. Game developers sometimes ask whether MT or MTPE can reduce their localization budget. This guide explains where MTPE works in game localization, where it doesn’t, and what the actual tradeoffs are.

Where MTPE Can Work in Game Localization

MTPE is most viable for content types where linguistic quality requirements are lower and volume is high: tooltip descriptions for common items with predictable patterns; UI labels with standard vocabulary (‘Settings’, ‘Volume’, ‘Graphics’); system messages with fixed structure; and high-volume, low-prominence text (loading tips, minor NPC incidental dialogue). MTPE benefits from Translation Memory — TM-matched segments (previously translated content that matches current content) are not sent to MT at all. For games with large TM coverage from previous versions, MTPE applies to the genuinely new segments that have no TM match. Neural MT quality has improved significantly for major European language pairs (EN-DE, EN-FR, EN-ES), making MTPE more viable for these pairs than for less common language pairs or Asian languages.

Where MTPE Fails in Game Localization

MTPE produces consistently inferior results for: narrative dialogue and character voice (MT cannot maintain character personality, speech register, or narrative tone — post-editing narrative MT often takes longer than direct translation); culturally specific content (Japanese cultural references, humor, and idiom translate poorly by MT and require significant rewriting); highly creative content (item names, ability names, marketing copy, faction descriptions); RTL languages (Arabic, Hebrew, Farsi — MT output requires extensive review for naturalness and cultural appropriateness); Asian language pairs (Japanese, Korean, Chinese — MT quality from/to Asian languages is significantly lower than for European pairs, particularly for Japanese which has extreme structural difference from English). For premium game content — especially games where localization quality is a market differentiator — MTPE produces a detectably different result than human translation.

The True Cost Comparison

MTPE is not always cheaper than human translation when total cost is correctly calculated. MT per-word rates are lower than human translation rates, but MTPE projects require: MT engine licensing or API costs; MTPE editor time (which is not zero — MT output requires editing, not just acceptance); higher LQA rates to catch MT errors that would not occur in human translation; and potential re-translation cost when MT output is too poor to be fixed efficiently (which happens regularly in narrative content). For content types where MTPE works well, cost savings of 20–40% vs. full human translation are achievable. For content types where MTPE fails, total project cost can exceed equivalent human translation cost when re-work is included. The decision requires content-type analysis, not a blanket MTPE-for-everything approach.

SandVox’s Approach to Machine Translation

SandVox uses machine translation selectively as a tool within a human-led workflow — not as a wholesale replacement for human translation. Our approach: MT is considered for specific content types (high-volume, lower-creativity UI strings) when it can demonstrably reduce cost without reducing quality. All MT output is reviewed by professional post-editors. We never apply MT to narrative dialogue, character voice content, or marketing copy. We are transparent with clients about when MT is used and deliver at the same quality standard as human translation regardless of whether MT was part of the pipeline. For clients who specifically want MTPE (to achieve budget targets on appropriate content), we scope the project with clear content-type boundaries between MT-appropriate and human-only content.

Frequently Asked Questions