Why Your Teleprompter Breaks When You Switch Languages

What Happens When You Switch Languages Mid-Recording?

You're recording a talking-head video for a bilingual audience. The script runs in English, then drops into Spanish for a client testimonial line, then back to English. On paper it reads fine. On camera, the moment you switch languages, the screen stalls. It doesn't scroll ahead, it doesn't scroll back - it just sits there, waiting for words it isn't listening for. You finish the Spanish line, come back to English, and the teleprompter is two sentences behind, or worse, highlighting a line you already read.

This isn't a glitch specific to one app. It's a shared limitation across every voice-following teleprompter, including Followscript, because of how the underlying speech recognition is built. Understanding why makes the fix obvious, and it isn't "buy a better app."

Why Does Voice Recognition Break When You Switch Languages?

The speech recognition built into browsers - the API almost every voice-following teleprompter is built on - takes a lang property that accepts exactly one language tag, like en-US or es-ES. That's according to MDN's own reference documentation. There's no built-in way to hand it two tags at once, or have it detect a change mid-sentence. The standard fix is to stop the recognizer and restart it with a different lang value. That's not something you can do gracefully mid-sentence, on camera.

That single-language design doesn't just mean the app "doesn't understand" the second language. It's worse than that. Soniox, a speech-recognition vendor, explains the mechanism in its own technical documentation. A system configured for one language "cannot return a word outside its configured vocabulary, so it produces the nearest [language]-shaped thing" instead of an honest error. Combine that with phonetically similar words across languages - English "see" and Spanish "sí" sound almost identical. Soniox puts it plainly: these sound-alike words "resolve the wrong way every time." Confidence stays high the whole time, because the acoustic match is good. In plain terms: switch languages, and the system doesn't fail loudly. It fails quietly and confidently - which is exactly why the teleprompter can drift to the wrong line without any obvious warning.

How Much Worse Does Recognition Get When You Mix Two Languages?

The numbers aren't a rounding error. Gladia's 2026 ASR guide states plainly that "monolingual ASR exhibits 30-50% higher WER on code-switched data compared to clean monolingual audio" - and that's the mild end. Deepgram's own benchmarking found OpenAI's Whisper Turbo-V3 model degrading "from 22.04% monolingual Word Error Rate to 58.67% WER on Bahasa Malay-English code-switched audio." For language pairs that sound less alike, it's worse still. Deepgram reports "WER exceeded 114%, mostly from massive insertion errors" - the transcript ends up longer and more wrong than what was actually said. Across several public benchmarks, Deepgram summarizes the overall pattern in one line: code-switched audio produces "1.5x to 11x higher error rates than monolingual baselines."

Apple's own research team calls this "a significant challenge" - intra-sentential switching, mixing languages inside one sentence rather than between takes. Their best fix so far brings a monolingual-trained system's error rate down from 34.4% to 15.3%. That's still nearly double a clean single-language result. And it needed dedicated engineering most teleprompter apps haven't built. Whatever the exact percentage on your specific script, the direction holds everywhere it's been measured: switch languages mid-sentence, and recognition gets meaningfully worse - not just "a little less accurate."

What Are Competing Teleprompters Doing About Language Switching?

Sayscroll launched publicly this month. It advertises support for "60+ languages" and claims "you can mix languages inside one script." That's the vendor's own claim, not something independently verified here. Worth testing on your actual script before you trust it for a take you can't reshoot.

Cuefyr is more direct about where the limit still sits. It's a desktop teleprompter built specifically around multilingual support - English, Urdu, and Hindi, processed on-device rather than sent to a server. Even so, its own feature list says "mid-sentence English and Urdu code-switching is on the roadmap." A product built from the ground up for multiple languages hasn't shipped true mid-sentence switching yet, either.

None of this means multilingual teleprompters are dishonest. It means "supports language X" and "handles switching between languages mid-sentence" are two different claims - and only one of them is common right now.

How Do You Script a Video That Uses Two Languages?

Most of the pain here isn't the teleprompter - it's a script that scatters short phrases in a second language throughout otherwise single-language sentences. If you plan exactly where you switch languages before you write the final draft, the recording itself gets far less stressful. Restructure it before you record:

1. Read through your draft and mark every phrase that isn't in your primary language - a greeting, a product name, a quoted testimonial, a sign-off.
2. Group scattered phrases into their own short section where possible, instead of one word mid-sentence. "Gracias, and see you next week" is harder to track than a full sentence in Spanish followed by a full sentence in English.
3. Pick a recognition language before you script, not after. If most of the video is English, script it so the Spanish lines sit at natural breaks - end of a sentence, start of a new paragraph - not buried inside an English clause.
4. Flag the switch points visually in your script - bold, a different color, a line break. You'll know exactly where tracking is likely to stall, instead of getting caught off guard on camera.
5. Do one dry run out loud before you hit record, watching how the screen reacts at each switch point, not just reading the words silently.

Should You Record Separate Takes Per Language, or Push Through in One?

There's no single right answer here - it depends on how often you switch languages in the script and what you're recording for.

If a script is genuinely bilingual - roughly equal parts of two languages, not a phrase here and there - separate takes stitched together in editing beat fighting the same switch point on retake after retake. Mostly one language with a few short phrases in another? One careful take is usually faster. Script those phrases at clean sentence breaks (see the previous section) instead of booking a second recording session.

What If You Only Need a Few Words in a Second Language?

You don't need to fully switch languages for one borrowed word. A single foreign brand name, a number, or a two-word quote doesn't usually justify restructuring your whole script or recording style. This is the same problem covered from the writing side in how to format a teleprompter script voice recognition can actually follow. Acronyms, numbers, and unfamiliar words all trip up recognition, and the same formatting habits help here too: spell things out, write numbers as words, keep unfamiliar terms short. One added wrinkle for a genuine second language - consider whether it needs saying at all in the voice-tracked take, or whether an on-screen caption afterward is cleaner.

How Do You Recover When Tracking Loses You Mid-Switch?

A voice-controlled teleprompter is built to match your voice against the whole script, not just the next few words - which is exactly what makes recovery possible here. If tracking stalls right after a language switch, don't stop the take. Say your next line in the language the recognizer is actually set to, at a normal pace and volume, and it typically resyncs the moment it hears something it can confidently match. Repeating the phrase that caused the problem, louder or slower, doesn't help. The recognizer isn't struggling to hear you - it's structurally unable to recognize that language at all in that session, so repetition just wastes the take.

Recording a full take that deliberately mixes languages? Know in advance where your teleprompter's reading line is likely to drift. A glance down then costs you a second, not the whole take.

What Should You Check Before You Hit Record?

A short pre-record checklist, once your script is written:

1. Every point where you switch languages is marked visually in the script.
2. You've decided, per section, single word or real switch. A single word: leave it, format it per the acronym/number rules. A real language switch: give it its own sentence or its own take.
3. Your teleprompter's recognition language is set to match whichever language makes up most of the current take.
4. You've done one dry run out loud, specifically watching the screen at each switch point.
5. You know what you'll say if tracking stalls: the next scripted line, said plainly, not a louder repeat of the line that broke it.

None of this requires new hardware or a different app - it's mostly about writing the script so the switch points are visible to you before they become invisible failures on screen. For more on getting recognition to follow the words you actually wrote, see the speech recognition topic and recording topic, or browse reading line placement for related fixes.

Followscript is a teleprompter that follows your voice - free in the browser, no install, no watermark. Switch languages between takes, and it tracks each one just as closely as a single-language script.