# Language mixing URL: /tts/concepts/language-mixing How Soniox Text-to-Speech follows language changes naturally within one utterance while using the selected language for overall delivery. ## Overview Real-world text often includes foreign names, international addresses, technical terminology, and phrases from multiple languages. Soniox Text-to-Speech follows these language changes naturally within one continuous utterance with no need to split the text or stitch audio together. You still select a **primary language** with the `language` field. That choice sets the overall delivery and a light accent bias for the voice, while embedded words and phrases from other languages are pronounced in a natural way. ## How language mixing works Pass mixed-language text in a single request or stream session, and set `language` to the main language of the text or to the language you want the speech to sound like. ```json { "model": "tts-rt-v2", "language": "en", "voice": "Adrian", "text": "Please deliver the package to Calle de Alcalá 45 in Madrid, attention Señora García." } ``` In this example, English is the primary language, so delivery follows English prosody. Spanish place names and titles are still pronounced naturally within the same utterance. ## Choosing the language The `language` field remains required. Use it for either: * **Main language of the text** — best when most of the content is in one language and only a few words or phrases come from others. * **Desired delivery language** — best when you want the voice to sound as if it belongs to that language community, even if the written text is mostly something else. Sending the same text with a different `language` changes how the whole utterance sounds. Compare the example above with: ```json { "model": "tts-rt-v2", "language": "es", "voice": "Adrian", "text": "Please deliver the package to Calle de Alcalá 45 in Madrid, attention Señora García." } ``` Here the text is still mostly English, but Spanish now drives the delivery, so the speaker sounds Hispanic — a light accent bias rather than a heavy accent. The same idea works in the other direction: Spanish text with `"language": "en"` leans toward an English-speaking delivery while still handling Spanish words naturally. Don't pick `language` based on a few embedded words. It biases the accent of the entire utterance, not just those words. Foreign names, addresses, and terms are handled correctly without changing `language`. Every [voice](/tts/concepts/voices) works with all [supported languages](/tts/concepts/supported-languages), so you can keep the same speaker while changing `language` or mixing languages in the input.