Can AI finally solve the Vietnamese translation problem?

Vietnamese has punished machine translation systems for years, and not for subtle reasons. Six tones that change a word's entire meaning, a writing system layered with diacritics that older software routinely mangled, and significant vocabulary and pronunciation differences between northern, central, and southern dialects have combined to make Vietnamese one of the languages where automated translation has lagged furthest behind more commonly modeled languages.

Newer large language models have narrowed that gap considerably, according to several bilingual reviewers who track translation quality closely, particularly on straightforward conversational text. Where the systems still struggle, reviewers say, is tone and register — distinguishing the deeply hierarchical pronoun system that shifts based on the relative age and social relationship of speakers, a distinction English simply doesn't require and that a mistranslation can render unintentionally rude or oddly formal.

The stakes are more than academic for families who rely on these tools daily. Second-generation Vietnamese Americans increasingly use translation apps to communicate with grandparents, read old family documents, or navigate a hospital visit for an elderly relative — situations where a tone-deaf translation isn't just imperfect, it can be actively confusing or hurtful at an already stressful moment.

Several of the reviewers interviewed described a cautious optimism rather than full confidence: the tools are meaningfully better than they were even two years ago, good enough for many everyday situations, but still not something they'd trust unsupervised for anything legally or medically significant. "I use it to write a text to my grandmother," one said. "I would not use it to translate a lease."