Short answer: don’t put multiple languages in a single TTS prompt. ElevenLabs auto‑detects language per request, so mixed‑language text in one generation is unreliable. Use separate generations per language and then join them.
How to do it properly
Choose a multilingual TTS model that supports all target languages (e.g., Eleven v3, Multilingual v2, Flash v2.5). See the current language support list. (help.elevenlabs.io)
If you use the website UI: the language is auto‑detected and you can’t force it; avoid mixing languages in one prompt. Generate each language segment separately. (help.elevenlabs.io)
If you use the API: make one request per language segment and set language_code to the correct ISO‑639‑1 code for that segment. This enforces language for the model and text normalization. Example body fields to look for: text, model_id, language_code. (elevenlabs.io)
To make stitched audio sound continuous, pass previous_text and next_text when generating adjacent segments; these fields are designed to help when concatenating multiple generations. (elevenlabs.io)
Pick appropriate voices: any voice can speak any supported language, but accent/pronunciation comes from the voice. For best results, use a voice trained in the target language (or a native accent clone) for each segment. (help.elevenlabs.io)
Key references
How to select language and accent (UI vs API; avoid multiple languages in a single prompt; language_code usage). (help.elevenlabs.io)
Create speech API (language_code, previous_text/next_text, pronunciation dictionaries). (elevenlabs.io)
I'm using the website UI.That's gonna be tricky, sometimes I need the audio sounds as it is as English word, like AI, Machine Learning. It is because the spelling in my native language or it's translation would be so weird.
ruby condor
somber pivot