Transparency · AI

AI & Translation Disclaimer

Last updated: 2026-05-15 Plain language, no fine print

01Summary

SpeechKey uses artificial intelligence to convert speech and text from one language to another in real time. AI is powerful but imperfect. Translations may contain errors, omissions, ambiguities or invented details ("hallucinations"). You should treat AI-generated translations as a helpful aid, not as a certified or legally binding translation.

Don't rely on AI alone for contracts, court proceedings, medical instructions, drug dosages, legal disclosures, safety-critical instructions, financial commitments, immigration documents, or any communication where a misunderstanding could cause harm.

02How AI translation works (and where it can fail)

Translation models predict the most likely sequence of words in the target language given the input. They are very good at common phrasing, business conversation and conversational tone — and they can fail at:

  • Rare or domain-specific terminology (legal clauses, niche medical terms, regional dialect).
  • Numbers and units — the model can mis-transcribe digits, drop currencies, or normalise units unexpectedly.
  • Negation and conditionals — "not later than Tuesday" vs. "by Tuesday" is a frequent source of subtle errors.
  • Idioms and humour — literal translations of culturally embedded phrasing.
  • Names, brands and places — homophones can be mis-resolved.
  • Audio with overlapping speakers, low signal, or strong accents.

03Possible error modes

  • Omissions — entire short phrases skipped, especially at audio boundaries.
  • Substitutions — one word replaced by a phonetically similar one.
  • Hallucinations — content inserted that was not actually said, often plausible-sounding.
  • Mistimed punctuation — sentence boundaries placed in the wrong location, changing meaning.
  • Provider failover artefacts — when adaptive routing switches providers mid-conversation, a tiny stretch of text may be re-translated or briefly duplicated.

04High-stakes contexts

Do not use SpeechKey as the sole source of truth in these contexts:

  • Legal proceedings, contracts and signatures.
  • Medical diagnosis, clinical advice or prescription instructions.
  • Safety-critical instructions (industrial, transport, emergency).
  • Financial agreements with material consequences.
  • Government, immigration or visa procedures requiring certified translations.

For these contexts, use a qualified human translator or interpreter. SpeechKey can help you understand the gist or prepare; it does not replace certification.

05Human verification — when and how

If a translation is going to influence a real-world decision:

  • Read the transcript back to the speaker in their own language and confirm.
  • For numbers and dates, repeat them digit by digit.
  • For contractual or medical content, request a written summary and have it verified by a qualified human.
  • When in doubt, ask: "Did I understand correctly that …?" — and let the other party correct you.

06Liability for decisions

You are responsible for decisions you make based on AI-generated translations. As stated in our Terms, we do not warrant the accuracy of translations and exclude liability to the maximum extent permitted by law.

07AI transparency & EU AI Act

  • SpeechKey is clearly labelled as an AI-generated translation service. Conversation participants should be informed when AI translation is in use.
  • We support EU AI Act transparency obligations: AI-generated content is identifiable as such within the app and on exports.
  • We maintain a current list of supported AI providers and their models. Customers may inspect or restrict provider routing at the account level.
  • We do not use customer audio or transcripts to train AI models.

08Feedback & corrections

Found a problematic translation? Tap the transcript line and select Report. Aggregated reports help us monitor quality without ever exposing your conversation content to us in raw form.

See also: Terms, Compliance overview.