AI video dubbing platform
Our team built a video dubbing platform that took English video through timed transcription, Spanish translation and replacement audio to a downloadable final video.
New public demo with sample data, inspired by the project.
Updated 2026-10-26
The challenge
A translated transcript still leaves you with a video in the wrong language. Voz AI needed to carry a recording all the way through to a watchable Spanish version, while keeping the words and their place in the video connected.
Our approach
We connected video intake, timestamped transcription, translation and dubbed audio generation. The final stage removed the original audio, combined the translated speech with the source video and prepared it for download. The platform used AWS services alongside open-source models trained and deployed on AWS.
Inside the experience
A visual guide to the published product scope.
Bring the original recording into a processing job.
Keep timestamped speech connected to its Spanish translation.
Generate translated speech and combine it with the original video.
Check the result and download the dubbed video.
Conceptual illustration · Based on public capabilities, not a private architecture diagram.
Work by our team before Rubix Labs. Voz AI focused on realistic video dubbing, with an English-to-Spanish workflow. The images and interactive demo on this page show a new public reconstruction using a licensed interview excerpt and a prepared Spanish dub.
Our team built the path from uploaded video to a downloadable dubbed version. The web experience included transcription-job history, speaker and timestamp grouping, playback-linked text and uncertain-word highlighting. AWS services and open-source models trained and deployed on AWS supported the wider processing workflow.
The result was a video people could watch in Spanish, rather than a transcript they had to use in a separate tool. Job history and playback-linked text gave the processing pipeline a usable web interface.
The engineering problem crosses several boundaries: speech recognition can mishear a word, translation can change the length of a sentence, and the replacement audio still has to fit the video. Keeping timestamps and review context together matters throughout. Voice cloning and lip syncing were planned for later iterations; they are not presented here as delivered features or included in this demo. We do not publish model accuracy or audio-quality benchmarks for this project.
Compare prepared English and Spanish versions, seek through the timed transcript and download the sample Spanish video. This demonstrates the workflow, not the original models or their quality.
Product capabilities
We only include capabilities that are already public. We do not share client data, private architecture, commercial terms, or internal performance metrics.
Related case study
We built an in-house OCR workflow that extracts driver license data, reduces manual entry, and gives operations teams a review step.
Rubix Labs
A few finishing touches.
