ElevenLabs Dubbing v1 vs v2: Test on a 24-Second Clip

👁 117 reads

ElevenLabs Dubbing translates video into another language and lays down a new audio track with voices that aim to sound like the original speakers. We tested it on our own fictional clip: 24 seconds in Spanish, two synthetic voices and background sounds, dubbed into English with models v1 and v2. Below are both results to watch and listen to, times, credits used, voice similarity measurements, and one failure you won’t read about in the pricing page.

How ElevenLabs Dubbing Works

The service recognizes speech in the file, separates speakers, translates the text, and generates a new track, leaving the original background (music, sounds). As of September 2026, there are two models:

• Dubbing v1 - older, cheaper, supports the same languages as Eleven v3. Only v1 works with Dubbing Studio, which allows manual editing of transcripts and regenerating individual fragments. The documentation states that Dubbing Studio is in maintenance mode and receives only critical fixes.
• Dubbing v2 - “end-to-end” model, default in API and in automatic dubbing in the app (described there as Alpha). 92 languages according to the pricing page, voice similarity slider (cloning strength 0-10, default 7).

Limits: in the app, files up to 1 GB and 180 minutes; via API up to 3 GB; in Dubbing Studio up to 1 GB and 45 minutes. Dubbing Studio is available from the Starter plan onwards.

Our Test: 24-Second Fictional Clip from Spanish to English

We didn’t want to dub someone else’s video or a real voice, so we assembled the material ourselves. The picture is six 4-second shots from our video generator ranking (coffee in a kitchen and a tram in the rain). We recorded the lines with the Eleven v4 model using two voices from the library: narrator George and female voice Jessica. The background - kitchen and rain with a tram bell - we generated in ElevenLabs Sound Effects. Everything in Spanish, no subtitles.

We tested it in Spanish dubbed into English. On September 29, 2026, we sent the same MP4 file via API to both models with parameters source_language=es and target_language=en. Dubbing returns only the audio track (FLAC), so we joined it with the picture in ffmpeg.

We read v2 credits from the account usage history after the second run, and calculated v1 from the difference between readings, because the dubbing endpoint doesn’t return cost in the header. We measured pauses using word timestamps from Scribe transcription.

Translation: v1 Correct, v2 Sounds Natural

Both models translated the text without factual errors. The difference is in style. The line “¿Cinco minutos? Yo tengo dos si el tranvía llega tarde, y siempre llega tarde cuando llueve” was rendered by v1 as “Five minutes? I only get two if the tram happens to be late, and it is always running behind schedule whenever it starts to rain.” Model v2 translated it as “Five minutes? I get two if the tram is late, and it’s always late when it rains.” The ending “justo a tiempo, por una vez” v1 translated as “right on time for once,” and v2 as “right on time for once” (run 1) and “pulls up right on time, for once” (run 2). In our opinion: v2 sounds like a narrator who understands the tone, v1 like a correct translation from a page.

Comparing with the Non-English Test

In a separate test run in Polish, one of two v2 runs returned gibberish (a stream of random words in multiple alphabets) and still charged credits. In this English test, both v2 runs completed successfully with clean, natural translations, though they differed slightly from each other. The Polish test showed v2 can fail unpredictably; the English test showed it can also vary between runs even when successful.

Practical takeaway: before adding more languages, download the source transcription (endpoint /transcript) and check that it makes sense. It takes a few seconds.

Does the Voice Stay the Same

We didn’t judge this by ear. We compared voice embeddings (WavLM speaker verification model, scale from -1 to 1) in the same time windows in the original and in the dub.

V1 preserved timbre slightly better than v2, and the difference from a different speaker is clear. Changing language always lowers this metric because the voice speaks different phonemes. Model v2 wins on rhythm, not timbre.

How Much ElevenLabs Dubbing Costs

Rates from the API pricing page checked September 28, 2026, excluding VAT. Charged per minute of source material, separately for each language.

In subscription plans you pay with credits: Starter at 6 USD gives 30,000 credits per month, Creator at 22 USD gives 121,000. With the usage from our test (5559 credits for 24 s), the Creator plan is enough for about 9 minutes of v2 dubbing per month. In v2, free accounts get a watermark automatically, paid accounts don’t. Plan details are in the article how much ElevenLabs costs.

How to Dub Step by Step

1. In the app, go to Dubbing, upload a file or paste a link, and set the source language manually (automatic detection is an additional risk).
2. Choose your target language, leave the number of speakers on automatic unless you know how many there are.
3. After completion, listen to the whole thing. In v1 you can make corrections in Dubbing Studio; in v2, transcript editing via API is only available in the Enterprise plan according to the documentation.
4. Before publishing, check proper names and numbers. In the API you can provide them in keyterms so the model doesn’t mangle them.

Via API, two commands are enough. We tested them on September 28, 2026 on a 9-second fragment with model_id=dubbing_v1:

curl -X POST https://api.elevenlabs.io/v1/dubbing/project 
  -H "xi-api-key: $ELEVENLABS_API_KEY" 
  -F file=@video.mp4 -F source_language=es 
  -F target_language=en -F model_id=dubbing_v2

# then every few seconds, until status = completed:
curl -H "xi-api-key: $ELEVENLABS_API_KEY" 
  https://api.elevenlabs.io/v1/dubbing/project/PROJECT_ID/language/LANGUAGE_ID

In the response, the outputs.lossless_audio field contains a link to the FLAC file valid for one hour. You join picture and sound yourself, e.g. ffmpeg -i video.mp4 -i dub.flac -map 0:v -map 1:a -c:v copy -c:a aac output.mp4.

What to Watch Out For

• Other people’s voices. Dubbing reproduces the timbre of speakers, so by uploading a recording of another person, you’re cloning their voice. Only do this with their consent or on material you have rights to.
• Labeling. Viewers should know they’re hearing a generated voice. General principles are described in AI dubbing guidelines.
• Quality control. Status “completed” doesn’t mean the result makes sense, as the Polish test’s first v2 run showed.

A full overview of ElevenLabs tools is in the guide ElevenLabs overview.

Frequently Asked Questions

Does ElevenLabs support dubbing into English?

Yes. English is supported by both models, v1 and v2. In our test, both translated a 24-second clip from Spanish correctly, and v2 preserved the original’s pauses with accuracy up to 0.56 s.

How much does ElevenLabs dubbing cost?

Via API, v1 costs 0.33-0.50 USD per minute, and v2 costs 2.20 USD, excluding VAT. Our 24-second clip used 1235 credits in v1 and 5559 in v2.

Does ElevenLabs dubbing preserve the original voice?

Approximately. In our measurement, voice similarity between original and dub was 0.75-0.97 in both models, versus 0.40-0.50 for a different speaker.

Can you do ElevenLabs dubbing for free?

On the free plan with 10,000 credits per month, yes, but with a watermark and no commercial license. With our v2 usage, that’s about 44 seconds of material.

How long does dubbing take?

Our 24-second clip was ready after 17.0-24.3 s in v2 and after 29.8 s in v1. For longer files, ElevenLabs recommends webhooks instead of polling for status.

Kinetyra

The link to ElevenLabs is a partner link: if you buy through it, QUANTSLANT earns a commission and you pay the same price. It has no influence on the content, including where the coverage is critical.