Translation, in speech and language AI, is the automatic conversion of text or speech from a source language into a target language by a computational model, with the goal of preserving meaning, tone, terminology and format. Often called machine translation (MT), it is one of the oldest problems in artificial intelligence and today relies mainly on neural networks trained on large parallel and multilingual corpora.
Modern systems fall into two broad families. Dedicated neural machine translation (NMT) models are trained specifically to translate, while large language models (LLMs) translate as one of many skills and can follow instructions about style, glossary terms or audience. Both share the same core idea: encode the meaning of the source and generate a fluent target sentence token by token.
How neural machine translation works
A neural translation model is typically an encoder-decoder transformer. The encoder reads the source sentence, split into subword tokens, and builds a contextual representation of every token. The decoder then generates the target sentence one token at a time, using cross-attention to look back at the relevant source positions at each step.
Decoding usually relies on beam search with a beam width of about 4 to 8, which keeps several candidate translations in parallel and selects the most probable complete sentence. Decoder-only LLMs work differently in form but similarly in effect: the source text and an instruction form the prompt, and the model continues with the translation.
| Approach | Period of dominance | Core mechanism | Typical weakness |
|---|---|---|---|
| Rule-based MT | 1950s to 1990s | Hand-written grammars and dictionaries | Brittle and costly to extend |
| Statistical MT | About 1990 to 2016 | Phrase tables learned from parallel text | Choppy word order |
| Neural MT | 2016 onward | Encoder-decoder networks with attention | Needs large parallel data |
| LLM translation | 2023 onward | Instruction-following language models | Cost, latency, occasional additions |
Text, speech and multimodal translation
Translation is not limited to written text. Speech translation converts spoken input into text or speech in another language, and it can be built either as a cascade of separate components or as a single end-to-end model.
| Pipeline | Components | Advantage | Limitation |
|---|---|---|---|
| Text to text | One MT model | Mature and accurate | Needs clean input text |
| Cascaded speech translation | Speech-to-text, MT, text-to-speech | Each part replaceable and inspectable | Errors compound across stages |
| End-to-end speech translation | One model from audio to text or audio | Lower latency, keeps prosody cues | Harder to debug and adapt |
| Simultaneous translation | Streaming model with a wait policy | Output begins before the speaker finishes | Quality drops with low delay |
| Image text translation | OCR plus MT, or a multimodal model | Translates signs and documents | Layout reconstruction is hard |
Dubbing pipelines add text-to-speech and sometimes voice cloning so that translated speech keeps the original speaker’s voice. That raises consent and misuse questions that a text-only system never faces.
Measuring translation quality
Translation quality is hard to measure because many different outputs can be correct. Automatic metrics compare the output with one or more human reference translations, or estimate quality without any reference at all.
| Metric | How it works | Strength | Weakness |
|---|---|---|---|
| BLEU | N-gram overlap with references | Fast, long history | Penalizes valid paraphrases |
| chrF | Character n-gram overlap | Better for rich morphology | Still surface-based |
| COMET | Neural model trained on human ratings | Correlates well with human judgment | Opaque, model-dependent |
| Quality estimation | Predicts quality without references | Works in production | Less reliable on rare domains |
| MQM human review | Experts annotate typed errors by severity | Most informative | Slow and costly |
For decisions that matter, automatic scores should be confirmed by human evaluation on in-domain samples. A system can gain several BLEU points while still mistranslating the few terms a legal or medical client cares about most.
Choosing and configuring a translation system
The right setup depends on volume, domain, language pair and risk. High-resource pairs such as English and Spanish reach near-professional fluency for general text, whereas low-resource languages with little parallel data still produce noticeably weaker output.
| Situation | Recommended approach | Reason |
|---|---|---|
| High-volume product listings | Dedicated NMT with glossary | Low cost and consistent terms |
| Marketing copy needing a brand voice | LLM with style instructions plus review | Controls tone and adaptation |
| Contracts and medical texts | MT plus professional post-editing | Errors carry legal or health risk |
| Low-resource language pair | Multilingual model, human review | Training data is scarce |
| Live meetings | Streaming speech translation | Latency matters more than polish |
Terminology control is the most common customization. It can be implemented through glossary constraints, fine-tuning on translation memories or, with LLMs, by placing the glossary and examples of approved translations in the prompt.
Limitations and typical errors
- Fluent but wrong output. Neural systems rarely produce broken grammar, so errors are easy to miss.
- Omissions and additions. Models may drop a negation or, in the case of LLMs, add content not in the source.
- Lost context. Sentence-level systems mistranslate pronouns, gender and terms defined earlier in a document.
- Bias. Gender-neutral source words are often rendered with stereotyped gender in the target language.
- Formatting damage. Tags, numbers, units and placeholders can be altered or reordered.
These risks explain why machine translation post-editing has become a standard professional workflow: the model drafts, and a qualified translator verifies meaning, terminology and formatting before publication.
Explained without jargon
Machine translation is like a very well-read assistant who has seen millions of texts in two languages side by side. When it reads a new sentence, it does not look words up one by one. It grasps the overall meaning and then writes a natural sentence in the other language.
It is fast and usually fluent, but it can still misunderstand an ambiguous word or a phrase that depends on context. For anything important, a person who knows both languages should check the result.
A typical use case
An online retailer sells 20,000 products and wants to launch in Germany, France and Poland. Translating every description by hand would take months and cost far more than the expected first-year revenue.
The team runs all descriptions through a neural translation model with a 600-term glossary of product names, sizes and material terms. A quality estimation model flags about 8 percent of segments as uncertain, and only those go to professional post-editors, together with all safety warnings and legal pages. The catalog launches within six weeks, and customer questions about mistranslated sizes remain rare because sizes and units were locked by the glossary rather than left to the model.