Three stages, three possible sources of error
The service first recognises speech, then translates the resulting text, then generates spoken output. A missed word in the first stage can affect the final sentence. Audio quality, accents, language pairs and context influence the result; there is no universal accuracy percentage for every conversation.
Make your first test representative
Try a short sentence with the microphone, language pair and participants you intend to use. Check specialist terms and ask listeners whether the result is understandable. For medical, legal, safety-critical or other consequential decisions, use an appropriately qualified human rather than relying on an automatic translation.