Speech recognition is most accurate when it knows what it's listening to. You can now choose the spoken language when you upload a private file, and recognition uses the best model for that language from the first second. Accuracy holds up on noisy audio, strong accents, regional dialects, and short clips, and names, numbers, and punctuation come through intact.

Auto-detect remains available. The choice affects speech recognition only; the summary language is set separately. If you mostly work in one language, set a default in Preferences.