An AI video clip is, most of the time, moving image. To become a communication piece it needs voice, captions and a soundtrack. That finishing is cheap and quick, and much of it runs on a GPU rented for a short time. This article shows the flow, from script to final file. (GenBrasil's clips and interface are in Portuguese; the same flow works for any language Whisper supports.)

The flow

  1. Script. Narration text, short, split by scene.
  2. Narration. A voice model turns the text into audio.
  3. Captions. Whisper transcribes the narration and produces an SRT file with timings.
  4. Assembly. ffmpeg joins clips, audio and captions.

1. Script per scene

Write the narration thinking in 5-second scenes: about 12 to 15 words per scene at a natural pace. If you want help, the script can come from a language model through the AI by token API.

2. Narration

Open voice models give good results, and GPUBrazil has 1-click templates for Chatterbox and F5-TTS. The cost per minute of audio is very low; the article Build your own ChatGPT: voice has the math.

A real person's voice

Cloning someone's voice requires explicit consent, preferably in writing. Use your own voice or a voice with documented authorization. See the responsible-use checklist.

3. Automatic captions with Whisper

Whisper is an open speech-recognition model with strong multilingual performance, Portuguese included. With the command-line tool:

whisper narration.wav --model large-v3 --language pt --output_format srt

The result is a narration.srt with segments and their timings. Review the text: proper names, acronyms and technical terms are where recognition errs most. To transcribe a lot of material, see Whisper large-v3 at scale.

4. Assembly with ffmpeg

First join the clips and add the narration; then burn the captions into the video (“burned-in” captions, which show up on any platform):

# 1) clips + narration
ffmpeg -f concat -safe 0 -i list.txt -i narration.wav -map 0:v -map 1:a -c:v copy -c:a aac -shortest piece.mp4

# 2) burned-in captions
ffmpeg -i piece.mp4 -vf "subtitles=narration.srt:force_style='FontSize=22,Outline=2'" -c:a copy piece_captioned.mp4

If the video is vertical (9:16), increase the caption's bottom margin so it doesn't sit under the platform's buttons.

Where the GPU really comes in

  • Whisper large-v3: runs on CPU, but is much faster on GPU. For one short video the difference is small; for dozens of pieces it is large.
  • Voice: open voice models need a GPU to generate quickly.
  • ffmpeg: doesn't need a GPU for this flow.

In short, one short hourly GPU session is enough: bring up the environment, generate narration and captions for the day's pieces and shut it down. The clips themselves keep being generated on GenBrasil.

Does the video already come with audio?

Some models generate audio along with the image, which changes the flow: in that case Whisper can generate the captions straight from the clip's audio. Check the model's behavior before assembling narration externally.

Frequently asked questions

Does Whisper work well in Portuguese?

Yes, large-v3 performs well in Portuguese. Review proper names, acronyms and technical terms, where it errs most.

Do I need a GPU to caption?

Not required, but Whisper large-v3 is much faster on a GPU, which matters at volume.

Can I clone my voice to narrate the videos?

Yes, with your own voice or with the person's documented authorization. Cloning other people's voices without consent is a legal risk.

Make your video on GenBrasil

Eight AI video models in one account, with the cost shown before you generate and payment in reais by Pix or card. The interface is in Portuguese. The Free plan needs no card.

Go to GenBrasil ↗

Rather build your own workflow? See the hourly GPU templates.

Read next: post-production on a GPU · voice: listen and speak