Speech-to-speech keeps the performance; multilingual TTS invents one

/glossary/speech-to-speech is a recording in, a new timbre out — timing and emphasis survive. ElevenLabs Voice Change is audio-to-audio at 5 credits. ElevenLabs Multilingual is text-to-audio at 6 credits. Do not feed a script to the changer.

The glossary already owns the definition. This page owns the fork in the catalog: which ElevenLabs row matches the file you actually have. If you have a take, convert it. If you have words, synthesise them. Treating those as the same button is how a good read gets thrown away and how a script gets stuffed into a converter that has nothing to convert.

The glossary term is a recording in

Speech-to-speech takes a recording of one person talking and re-renders it in a different voice, keeping the original performance intact. That sentence is the whole product boundary. Timing, emphasis, pauses, the laugh in the middle of a line survive because you acted them. The converter swaps timbre. It does not invent a delivery.

Text-to-speech is the other boundary: written text in, a synthetic voice out. The engine has to guess where the stress goes, how long to hold a beat, whether the line is a joke. Plain text does not contain a performance. Two takes of the same script are two invented readings. That is a feature when you are casting. It is a loss when you already directed the take.

Do not rewrite the glossary as a prompt knob. There is no slider that turns a paragraph into a performance. The input type is the knob: recording versus script.

The ceiling is the source. Conversion carries over what it hears. A mumbled line stays mumbled. A flat read becomes a flat read in a different timbre. If the complaint was the performance, conversion is the wrong tool.

ElevenLabs Voice Change is audio-to-audio at 5 credits

ElevenLabs Voice Change is the catalog row for that conversion. Content type audio. Category audio-to-audio. 5 credits. Catalog audio true. No resolution. No duration enum. You start with a recording. You get a recording. You do not get a face.

Use it when the words and the timing are keepers and the voice is not. Placeholder narration swapped for a roster voice. A temp read swapped for a cloned timbre you already have consent to use. Off-camera dialogue that should not be re-typed.

It will not invent a presenter or a plate. It will not lip-sync a mouth. If we see the face, a converted file laid on a closed mouth is still a closed mouth. Send the new track into a lipsync job, or convert audio that was always meant to live off camera.

Change the voice in a video is the picture-side cousin: replace the bed on a clip. Voice Change is the audio-to-audio engine. Trim before you convert. Keep the original. Conversion is a generate, not an edit.

ElevenLabs Multilingual is text-to-audio at 6 credits

ElevenLabs Multilingual is multilingual text-to-speech powered by ElevenLabs Multilingual v2. Content type audio. Category text-to-audio. 6 credits. No still. No duration enum. You start with a script. You get a performance the model invented.

That is the correct row when there is no take: a new VO, a language pass from copy, a line nobody has spoken yet. Punctuation is prosody. Write numbers the way they should be read. Audition more than one take; the output is stochastic.

It is the wrong row when the pause before the punchline already exists as a file. Re-typing that line into Multilingual throws the pause away and asks the engine to rediscover it. Sometimes it will. Often it will not. The 6-credit sticker is not cheaper conversion. It is a different input.

AI text to speech is the tool shape for that family: engines, not just a voice dropdown. Multilingual is one catalog row on that shape. Voice Change is not on that shape. Do not open TTS because the converter felt obscure.

Do not feed a script to the changer

The changer has nothing to convert if you paste words. A script is not a recording. Feeding copy into an audio-to-audio row is a mislabel, the same as uploading silence and hoping for a read.

The intake is five checks:

1. Is the delivery actually right? If not, recut or re-record. Then convert.
2. Is the vocal isolated? A music bed baked into the take comes along for the ride.
3. Is there one speaker? Two voices overlapping are not one target.
4. Are the lips on camera? Conversion does not re-time a mouth.
5. Do you have the right to put those words in that voice?

Four yeses and a clean consent answer means Voice Change. A script with no file means Multilingual. A close-up talking head means lipsync after, or a reshoot.

Do not use Voice Change to "fix" a bad read. It changes who is speaking, not how well the line was delivered. Do not use Multilingual to keep a performance you never recorded. Same vendor family. Different content type. Pick the row that matches the input you actually have.

FAQ

Can I paste a script into ElevenLabs Voice Change?

No. Voice Change is audio-to-audio at 5 credits. It needs a recording. A script belongs on ElevenLabs Multilingual at 6 credits, text-to-audio.

Does conversion rewrite the words?

It is not supposed to. Speech-to-speech preserves speech content and swaps timbre. If you need different words, record them or run text-to-speech. Do not hope the changer will ad-lib.

How is this different from the glossary page?

The glossary defines the term. This page is the catalog fork: Voice Change versus Multilingual, 5 credits versus 6, recording versus script. Do not treat the term as a prompt knob.

If the mouth is on camera, which row do I pick?

Neither row writes a talking-head video. Convert or synthesise the audio, then lipsync the plate, or shoot a real person. Laying a new timbre on a closed mouth is still a closed mouth.