A bilingual voiceover can sound natural until it reaches a name, a borrowed phrase, or the first sentence in another language. Then the accent shifts, a word changes, or the voice seems to become a different person. Mixed-language text to speech needs a boundary test before you generate the full script.
Download the language-boundary listening sheet · Audition a bilingual line in Voice Lab
Separate four problems that sound similar
First identify what actually failed. A name pronounced with the surrounding language's rules is a pronunciation problem. A sentence spoken in an unintended language is a language-selection problem. A correct sentence with an unsuitable regional accent is a voice-selection problem. A change in vocal tone or apparent speaker is a continuity problem.
These can happen together, but they need different responses. Rewriting a name phonetically may help that name while leaving the overall accent wrong. Switching to another voice may solve the accent while breaking continuity with the rest of a campaign.
| Symptom | Check first | Avoid as the first fix |
|---|---|---|
| One name sounds wrong | Approved pronunciation and model controls | Rewriting the entire script |
| Language changes unexpectedly | Supported languages and language selection | More emotional direction |
| Accent varies between phrases | Voice's suitability for both languages | Assuming one “multilingual” label guarantees every accent |
| Voice identity changes | Same model, voice and settings | Mixing unrelated voices without an editorial reason |
| Join sounds abrupt | Phrase boundary, level and cadence | Cutting through a consonant |
Write the failure in one sentence and retain the exact input that produced it. A short reproducible line gives you a better comparison than listening to several long, slightly different scripts.
Check the model before inventing new syntax
Language mixing is model-specific. ElevenLabs explains that a voice may retain its native accent or drift when used across languages. Its language-selection help also cautions about mixed-language prompts in that documented workflow. Other systems take a different approach: Soniox documents mixed-language input with a primary-language setting.
That difference matters. Do not paste an API field, SSML tag or provider-specific control into an ordinary speech box and assume it will work. Some systems may read the markup aloud. Check the chosen interface and model, not just the provider's general marketing page.
Choose the audience's intended language and regional delivery deliberately. Spanish for a particular audience is not a single universal accent, and an accent should not be treated as an error merely because it differs from yours. The review question is whether it fits the brief and remains understandable and consistent.
Build a short boundary test
Use this fictional café announcement to expose a switch without committing a full production script. The words are an exercise; replace the business details with verified information before using it publicly.
The passage contains an English opening, a Spanish product phrase, a complete Spanish sentence and a return to English. Record what should happen at each boundary. “Casa Clara” is a proper name, while “Pide el tuyo en la barra” is a full language change. Those are different tests.
Listen first without reading the script. Then follow the text word by word. Does the product phrase sound intentional? Does the Spanish sentence remain clear? Does the return to English change the apparent speaker? Ask a fluent reviewer to check meaning and naturalness rather than using your own confidence as the only standard.
Compare one-pass and phrase-based production
Candidate A uses the complete mixed-language passage with a model documented to handle that use. Keep the selected voice and ordinary delivery settings stable. Candidate B separates the script into complete phrases, using appropriate supported language settings where available. Preserve the same suitable voice when possible.
Do not split “pan de canela” into separate generated words. The phrase has its own rhythm, and isolated words can sound pasted together. Likewise, avoid joining files in the middle of “Welcome to Casa Clara.” Choose a natural pause or a complete sentence boundary.
Compare the candidates at similar playback levels. A louder voice can seem clearer without being more accurate. Judge pronunciation, continuity and pacing separately, then choose the method that performs best on your actual line. This is your audition, not a universal benchmark for a provider.
Keep a language and pronunciation ledger
| Script element | Intended treatment | Reviewer note |
|---|---|---|
| Casa Clara | Approved business-name pronunciation | Confirm with the owner or supplied reference |
| pan de canela | Spanish phrase inside English context | Keep all three words together |
| Pide el tuyo en la barra | Complete Spanish sentence | Check natural delivery and intended meaning |
| Return to English | Same overall speaker and energy | Avoid a sudden change in pace or tone |
For unfamiliar names, request a short pronunciation reference from someone who knows the intended form. Where a model supports a pronunciation dictionary or alias, use its documented method. Otherwise, a carefully tested spelling workaround may help, but keep the correct written form for captions and published copy.
Do not translate names or change meaning just to make the synthesizer comfortable. The voice file serves the script. If the line cannot be delivered accurately with the selected voice, choose a better-suited voice or record that phrase with an authorized speaker.
Join phrases without making the edit audible
Bring the accepted phrases into an audio or video editor. Place them in script order and listen through each join. Keep enough room for the preceding word to finish. A short pause can make a language switch intentional; too much silence can make it sound like a new segment or speaker.
Match perceived level and general tone. Use a small fade at a clean boundary if needed to avoid a click, but do not fade away the beginning of a consonant. Avoid stretching a phrase heavily just to fit a timeline; first revise unnecessary wording or adjust the edit around it.
Listen in context with the music bed, then solo the voice again. Background sound can hide a bad join during production, only for it to become obvious on headphones. The voiceover timing guide can help when one language takes longer than the available shot.
Review the exported voice and its captions
Check the final audio against the approved script, including names, numbers and any offer details. Generate or edit captions from the correct written copy, then align them to the accepted voice. Do not publish phonetic spellings that were used only to influence speech.
Have a fluent listener review the complete export, not just the test phrase. A successful audition does not prove the rest of a longer bilingual script is correct. Keep the language ledger with the accepted files so future revisions use the same names, voice decisions and regional brief.
If a revision changes only one sentence, recheck the surrounding joins and overall speaker continuity. Small replacements can sound brighter, faster or more expressive than an older take even when the voice setting has the same name.
How QuestStudio helps
Use Voice Lab to audition a short line with a model that supports the languages you need. The language control appears for models that support it; check the current selection before generating. Keep accepted phrases and pronunciation notes together, then assemble and review the final voiceover in your editor.
Frequently asked questions
Does multilingual mean perfect language switching?
No. Supported languages, voice training, context and interface controls all affect the result. Test the actual boundary you need to use.
Should I put language tags in the speech box?
Only when the selected model documents that syntax for that field. Otherwise the tags may be ignored or spoken aloud.
Can I use one voice for English and Spanish?
Potentially, if the model and voice support both appropriately. Audition pronunciation, regional delivery and speaker continuity before committing the script.
Is phonetic spelling okay in captions?
Keep captions in their correct written form. A phonetic workaround belongs in the speech-production input, not automatically in the published text.
What if the Spanish phrase is longer than the shot?
Revise the script or timing first. Avoid cutting words or excessively speeding up the phrase just to force it into the old duration.
Approve the switch before the full script
Finish the short test and decide whether one-pass generation or phrase-based production serves the line. Then scale the accepted method in Voice Lab, with a fluent listener checking the final result.
The worked example is a proposed creative exercise, not a measured model benchmark. The hero is an original AI-generated editorial illustration.