Case study

Arabic generative video pipeline

Multimodal generation with native Arabic audio and burned-in subtitles

Sanitized production case studyRole · POC lead and integration spec author

A generation pipeline that turns a brief into short cinematic video with native Arabic narration and on-screen subtitles — combining video, image, text, and speech models behind one orchestrated workflow.

Problem

Public-sector communications needed short-form video in Arabic, but off-the-shelf text-to-video produced motion artifacts and hallucinated or mismatched audio, and handled Arabic narration and right-to-left subtitles poorly.

My responsibility

Built the POC and delivered the integration specification: model selection, the generation pattern that fixed motion and audio quality, Arabic TTS and RTL subtitle burn-in, and the contracts (WebSocket, storage, worker model, session state) for integrating it into the wider platform.

Key decisions

Adopt an Acts + Extensions generation pattern — compose longer sequences from controlled segments — to resolve the motion-quality and audio-hallucination problems single-shot generation produced.
Generate Arabic narration with a dedicated speech service rather than the video model's audio, then align and burn in RTL subtitles for accessibility.
Combine specialized models per modality (video, image, text, speech) instead of forcing one model to do everything.
Specify the full integration contract up front — WebSocket events, blob storage layout, worker model, session state machine — so the POC could become a platform feature, not a dead-end demo.

Architecture

Input
Orchestration
LLM / generative
Deterministic
Output

A brief is expanded into a shot or act plan, each act is generated and extended for motion stability, narrated by an Arabic speech service, then composited with burned-in RTL subtitles and delivered via a worker-backed, WebSocket-observable session.

Stack

Azure OpenAI
Veo (video)
Gemini (image)
Azure AI Speech
Python
WebSockets
Blob storage

Outcome

· POC producing cinematic short video with native Arabic audio and burned-in RTL subtitles.
· Acts + Extensions pattern that measurably reduced motion artifacts and audio hallucination versus single-shot generation.
· Complete integration spec (WebSocket contracts, storage layout, worker model, session state) for platform adoption.

Known limitations

· POC-stage: built to prove feasibility and define integration, not yet hardened for high-volume production.
· Generation cost and latency are significant; suited to curated output, not real-time.
· Model availability and terms for the underlying video and image models can change.
· Client details and specific figures are sanitized and available under NDA during a technical interview.

Sanitized details available during technical interview.