Chain Gang All Stars

How Chain Gang All Stars Actually Works Under the Hood

Chain Gang All Stars is a vocal cloning and AI cover generation service. You upload a reference audio file of a singer, feed it a target instrumental or vocal stem, and the system transplants the voice onto the new track. It has been around since early 2023 and went mainstream when AI Drake and The Weeknd tracks flooded the internet. The platform itself is web-based at chaingangallstars.com, and they offer both a free tier with limited daily generations and paid credits for higher quality output and batch processing.

The pipeline isn't magic. It runs on a combination of RVC (Retrieval-based Voice Conversion) models and UVR5 for stem separation. When you upload an audio file, the backend first isolates vocals using a model like MDX-Net or Demucs. Then it converts the target vocals into the reference voice using an RVC v2 model trained on your sample. The result is then mixed back with the instrumental. You go to the website and sign up with an email. The free tier gives you a handful of credits per day. Each generation costs somewhere between 5 to 30 credits depending on the length and whether you're doing a solo or group pass. Paid plans start around $10 per month for a credit bundle. There is no desktop download. It is entirely browser-based. Some people run local copies of RVC and stitch the workflow together, but the hosted service handles the GPU compute on their end so you don't need your own hardware. I got tripped up early on because I kept uploading full mixed songs and wondering why the results sounded thin. The issue is that the system works best when you feed it a clean a cappella or a well-separated vocal stem. If you upload a fully mixed track, the model tries to isolate the voice from a dense mix and the transcription gets muddy. I switched to running the audio through UVR5 first, pulling out the vocal stem, then feeding that into Chain Gang All Stars. The turnaround time dropped from about 4 minutes per generation to under 90 seconds, and the quality jumped noticeably.

Another thing nobody tells you about the free tier: the sample rate gets downsampled aggressively. You end up with something that sounds like it was recorded through a doorbell. If you're generating for anything other than casual sharing, bump to a paid plan. The difference between the free and paid output is not marginal. It is the gap between "this sounds like a meme" and "this could actually pass in a demo." I lost three days wrestling with the free tier before I just paid for a month and accepted it as the cost of doing this seriously. The interface itself is functional but bare. You pick your reference voice or upload a custom one, paste a YouTube URL or upload audio for the target track, select the model quality, and hit generate. That is basically it. There is no waveform editor, no fine-tuning of pitch drift correction beyond a slider, and no post-processing chain. What you get is what you get. The biggest bottleneck I ran into was reference audio length. The default minimum is about 10 seconds of clean singing, and the model performs worst when the reference contains heavy reverb or background instrumentation. I solved this by recording a dry vocal take into my USB mic in a closet full of blankets, ran it through a noise gate, and used that as the reference instead of trying to scrape a studio recording. The dryness actually helps the model lock onto the timbre more precisely. Wet vocals introduce formant confusion that the RVC model struggles to separate from the target voice.

There are legitimate limitations here. The system struggles with non-sung vocals like spoken word or rap at high flows. It also has a hard time with voice types that are far outside its training distribution, which means extreme bass voices or whistle-register soprano will often sound wrong even when everything else is clean. And if your target track has a lot of vibrato or melisma, the pitch correction can slip and you get that telltale robotic warble that every AI cover listener recognizes instantly. If you need more control than this hosted service offers, the open-source RVC project on GitHub is the alternative. It requires a GPU with at least 8GB VRAM, takes a few hours to set up if you've never used Python environments, and you train your own model rather than uploading a clip. But you get direct access to inferencing parameters like filter radius, pitch shift, and index ratio. Chain Gang All Stars abstracts all of that away, which is why it is easier to use but harder to fine-tune.