GPU time is a small line in this decision. What decides it is whether you can make translated speech fit the video, and whether your voice model's licence allows commercial use.
Should you build your own dubbing pipeline on GPUs, or pay a dubbing API per minute? The question comes up as soon as a product team prices video translation at real volumes.
AI dubbing is a chain of four stages
AI dubbing turns speech in one language into speech in another, in four steps:
- Transcription: speech becomes text, with a timestamp for every word.
- Translation: the text moves into the target language.
- Voice synthesis, also called text to speech (TTS): a synthetic voice reads the translated text.
- Alignment: the new speech is fitted into the timing of the original video.
A dubbing API sells all four as one price per minute. ElevenLabs charges $2.20 per source minute for Dubbing v2, which it describes as end-to-end dubbing that preserves voice and emotion. Its older Dubbing v1 costs $0.50 per minute without a watermark. A build lets you buy or host each stage separately, so compare stage by stage.
AI dubbing cost per finished minute, stage by stage
A finished minute is the fair unit, because APIs bill per minute or per character and GPUs bill per second. To turn GPU time into minutes, use the real-time factor (RTF): the seconds of GPU time that one second of audio takes. An RTF of 0.15 means one minute of audio needs 9 seconds of GPU time.
GPU seconds per finished minute = 60 × RTF
cost per finished minute = GPU seconds × your price per GPU second
Our running example is hypothetical: 1,000 minutes a month of English training videos dubbed into German. We assume 150 spoken words per minute and 6 characters per word, so 900 characters. German text runs longer: the W3C cites an IBM guideline of 130% for translated strings over 70 characters. That leaves 1,170 characters to speak in each finished minute.
| Stage | Bought, price read 30 October 2026 | Bought, per finished minute | Built, GPU seconds per minute |
|---|---|---|---|
| Transcription | $0.22 per audio hour, ElevenLabs Scribe v2 | $0.0037 | 1.3, faster-whisper |
| Translation | Qwen3.8 27B on our API, $0.40 in and $2.40 out per 1M tokens | $0.0008 | none, it stays on the API |
| Voice synthesis | $0.08 per 1,000 characters, ElevenLabs multilingual models | $0.0936 | 8.8, F5-TTS |
| Alignment | not sold on its own | inside the bundle | engineering time |
| Total | about $0.10 | 10.1 |
The GPU figures come from published benchmarks. faster-whisper, an optimised version of OpenAI's Whisper speech model, transcribed 13 minutes of audio in 17 seconds. It used Whisper large-v2 on an NVIDIA RTX 3070 Ti, a consumer GPU, with 8 audio files processed together. F5-TTS, an open voice model, reached an RTF of 0.1467 on one NVIDIA L20, a data-centre GPU.
transcription: 17 s ÷ 13 min = 1.31 GPU s per minute
break-even: 3,600 ÷ 1.31 ÷ 60 = 45.9 audio hours per GPU hour
45.9 × $0.22 = $10.09 of API charges per GPU hour
So self-hosted transcription saves money only while an hour of GPU time costs less than $10.09 and stays busy.
Translation costs $0.0008 a minute on Qwen3.8 27B Instant, the variant with reasoning switched off, billed at the Qwen3.8 27B price. We assume 1.3 tokens per word, and German output about 30% longer:
tokens in = 150 words × 1.3 + 300 instruction = 495
tokens out = 195 × 1.3 = 254
cost = 495 × $0.40 ÷ 1M + 254 × $2.40 ÷ 1M = $0.0008
The bought voice price of $0.08 per 1,000 characters applies to any voice in your library, including a cloned one. Separate stages can keep the speaker's voice too, but you wire them together.
Across 1,000 minutes, the separate stages cost $98 a month and Dubbing v2 costs $2,200. Built, the GPU stages need 2.8 GPU hours a month, about $6.43 on an A100. Price basis: the on-demand price on our GPU compute page, read 30 October 2026.
GPU hours per month = 10.1 s × 1,000 min ÷ 3,600 = 2.8
The benchmarks ran on other GPUs, so measure your own RTF before you rely on it. The example favours the build: it leaves out quality, idle GPU time and the engineering hours a build needs.
For sizing Whisper itself, see our guide to deploying Whisper large-v3 behind a GPU API.
Voice synthesis is where the API is hard to beat
On compute alone, self-hosted voice synthesis wins by a wide margin. One GPU hour at an RTF of 0.1467 produces 409 finished minutes. At the bought voice price, the API would charge $38.28 for the same audio.
finished minutes per GPU hour = 3,600 ÷ (60 × 0.1467) = 409
API charge for that audio = 409 × $0.0936 = $38.28
So building is cheaper on compute whenever an hour of GPU time costs you less than $38.28 and the GPU stays busy.
Two things stop this arithmetic from deciding. The first is the licence. Read the model card before you build on a checkpoint, which is a set of trained weights:
| Open voice model | What its card or repository states | Commercial use |
|---|---|---|
| F5-TTS | Pre-trained models under CC-BY-NC, a non-commercial licence, because of Emilia, the speech dataset it was trained on | No |
| XTTS-v2 | Clones a voice from a 6-second clip, 17 languages including German, Coqui Public Model License, Coqui's own model licence | Read the licence text |
| Parler-TTS Large v1 | 2.2 billion parameters, Apache 2.0, model card lists English | Yes, English only |
The second is quality. A rate card cannot tell you whether listeners accept a voice. Naturalness and speaker similarity need a named, dated evaluation, such as a blind listening test with your own script and target language. The F5-TTS paper reports an RTF of 0.15 and strong zero-shot results, meaning voices it never trained on, from a short sample. It does not test your speakers.
Alignment is the stage that breaks builds
Alignment fits the translated speech into the time the original speaker took. German runs about 30% longer, as above, so a pipeline that reads the translation as written overruns the scene. Speeding up the audio to fit makes the voice sound rushed.
A build therefore controls length before synthesis. Transcription gives each sentence a start and end time. Translation then gets a character budget per sentence, and synthesis gets a target speed. F5-TTS lists speed control among its features. The budget goes into the translation request.
We are Lyceum, a sovereign European AI inference provider for open-source models: built and run in Europe, with our serverless models in Paris and Finland. This is the request on our OpenAI-compatible API:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["LYCEUM_API_KEY"],
base_url="https://api.lyceum.technology/openai/v1",
)
segment = {"text": "Open the settings and choose a new password.", "max_chars": 52}
response = client.chat.completions.create(
model="qwen/qwen3.8-27b-instant",
messages=[
{"role": "system", "content": "Translate into German for dubbing. Stay within the character limit."},
{"role": "user", "content": f"Limit: {segment['max_chars']} characters.\n{segment['text']}"},
],
)
print(response.choices[0].message.content)
Check every result against its budget and send overruns back for a shorter version. That loop, plus pauses and speaker changes, is the work the bundle price pays for. Price it in engineering weeks, not GPU hours.
Voice rights and consent sit outside the cost model
A synthetic voice that resembles an identifiable person raises rights and consent questions, whatever it costs to generate. We flag them here and do not resolve them. Take them to your legal team before the first cloned voice, not after launch.
Under Article 9 of the GDPR, biometric data used to uniquely identify a person is a special category. Its processing is prohibited unless an exception applies, such as explicit consent for specified purposes.
Article 50 of the EU AI Act asks providers, who make a system that generates synthetic audio, to mark its output in a machine-readable format. Deployers, who use such a system, must disclose that a deep fake was artificially generated. A deep fake is generated content that resembles existing persons or events and would falsely appear authentic. A dubbed voice that imitates the original speaker may fall under it, so consent belongs in the product design.
Splitting the chain between build and buy
At low volume, buy the bundle. At 100 minutes a month, Dubbing v2 costs $210 more than the separate stages. Compare that gap with what building alignment would cost your team.
Unbundle when the monthly gap pays for building and running alignment. In our example the gap is $2,102 a month at 1,000 minutes. The split we suggest then looks like this:
- Transcription and translation run wherever is easiest.
- The voice stays bought until an open model passes your listening test under a commercial licence.
- Alignment is built, because that is where the bundle's price goes.
The translation stage runs on our API, and prompts and outputs are processed, not stored, and never used for training. The GPU stages run as a batch job, which we call a Run. A Run is billed only for active execution time, while a GPU virtual machine bills its whole uptime.
lyceum docker run your-team/dub-batch -m gpu.a100 -c "python dub.py"
Spot GPUs are spare capacity at about 50% less, and we can reclaim them at any time. Process one video per run, so a reclaimed run loses one video that you run again.
Start with one 10-minute video. Time each stage, note its RTF and price it with the formula above. Then decide stage by stage on our GPU compute page.
Frequently asked questions
Is a dubbing API billed on the source minute or the dubbed minute?
ElevenLabs bills dubbing per source audio minute, so a German track that runs longer than the English original costs the same. Text to speech is billed per character, so longer German text costs more there.
What does the cheapest automatic dubbing tier include?
ElevenLabs Dubbing v1 costs $0.33 a minute with a watermark and $0.50 without, read on 30 October 2026. It detects speakers automatically and supports 29 languages, against 92 for Dubbing v2.