Vibechanger

Vibechanger is my shot at a video voice-style transformation pipeline.

It takes an input video, transcribes its audio, rewrites the transcript to a target tone, synthesizes new speech with CosyVoice3 using the speaker's own reference sample, and runs MuseTalk to produce a lip-synced output video.

How do you do that

Whisper, CosyVoice3, MuseTalk

Repo Structure

• jobs/video_ai/: main worker image and runtime pipeline.
• jobs/hf_downloader/: model downloader job for Hugging Face / direct URLs.
• scripts/run_video_model.sh: Cloud Run Job execution + log tailing wrapper.
• terraform/: infra for jobs, buckets, IAM, and related resources.
• ui/: local web UI for uploads/triggering job workflows.
• local/: local helper scripts and sample inputs/outputs.

Main Commands

Show all commands:

make help

Auth/project setup:

make auth-login
make auth-project PROJECT_ID=<your-project-id>

Build and publish images:

make docker-publish

Terraform workflow:

make terraform-init
make terraform-plan
make terraform-apply

Run the local UI to record your video

make ui

Or Use local input to give it a spin

make run

Download default model set to the models bucket:

make hf-download

Inputs and Outputs

• Local default input video: local/input/input.mp4
• Local output directory: local/output/
• Cloud job data paths are mounted under /data in the container.
• Model assets are mounted/read from /models in the container.

Challenges

I use Google Cloud Run jobs, its ephemeral and just easier to terraform. Models are backed by the cloud storage which can be loaded to the containers on runtime. Bootup time and Image size is still a bit harsh, separating the steps would need a messaging layer between them.