What AI is used for music covers?

The AI Cover Explosion: Flooded Uploads, Invisible to Listeners

Every day, roughly 50,000 fully AI-generated tracks hit Deezer's servers — 34% of all daily uploads — yet those same tracks account for only about 0.5% of total streams, according to Deezer/Ipsos research from late 2025. The floodgates are wide open, but listeners are barely drinking.

That disconnect defines the modern AI cover landscape. The tools are cheap, fast, and increasingly good: custom voice model generation has dropped from roughly 90 minutes three years ago to under five minutes today, with some platforms offering instant cloning from just 10–30 seconds of audio. Meanwhile, Deezer/Ipsos survey data across 9,000 listeners in eight countries found that 97% cannot reliably identify AI-generated music — meaning most covers slip past audiences undetected.

So why isn't every creator pumping out covers? Because the hard part isn't access anymore — it's choosing well. The space splits into two very different paths:

• Open-source pipelines like AICoverGen, built on RVC v2 voice conversion, MDXNET vocal separation, HuBERT feature extraction, and rmvpe pitch detection — powerful, but requiring Python, GPU resources, and technical patience.
• Commercial platforms like Jammable (60,000+ pre-made voices), Kits AI (licensed artist models from $10/month), ElevenLabs, and Soundverse Voice Swap — faster to use, but with opaque underlying models and vendor-reported stats that lack independent verification.
• Legal exposure that varies wildly: models trained on copyrighted recordings can produce outputs that infringe, and platform policies diverge sharply — Deezer labels AI music, Spotify removed 75 million spam tracks without labeling AI, and YouTube has no published policy at all.

This is the same evaluation problem we see across the AI ecosystem at SkillzBench: a flood of tools and components where quality, safety, and legitimacy are nearly impossible to judge from the outside. Whether you're picking a voice model or an agent skill, the gap between "works" and "works well, legally, and reliably" is enormous — and growing.

The stakes are real. A cover made with the wrong voice model can get your track de-prioritized, removed, or worse. And with 70% of streaming users believing fully AI-generated music threatens musicians' livelihoods, the audience tolerance for sloppy, unethical AI covers is thin. The tools exist; the challenge is picking the right one — and that's exactly what the rest of this guide untangles.

The Core Answer: RVC v2 and the Open-Source Cover Stack

Ask a dozen AI cover creators what powers their tracks, and you'll get the same answer: it isn't a song generator at all. AI covers are built with voice conversion — a model that takes an existing vocal performance and re-sings it in someone else's voice, while the original instrumental stays untouched.

The heart of that pipeline is RVC v2 (Retrieval-based Voice Conversion), the open-source model that has become the de facto standard for AI covers. Sonarworks identifies RVC as the most popular open-source option for voice conversion in music, prized for its low latency and vocal fidelity, and notes that it succeeded the earlier So-VITS SVC era. RVC voice models are trained on recordings of a target voice, then applied to a source vocal — which is why a cover sounds like the artist "actually singing" the song rather than speaking over it.

But RVC alone doesn't make a cover. The open-source AICoverGen pipeline documents the full stack of specialized models that work together:

• RVC v2 voice models (.pth files plus optional .index files) that perform the actual vocal conversion
• MDXNET separation models, which isolate the vocals from the instrumental so only the voice gets converted
• HuBERT, a base model used for audio feature extraction before conversion
• rmvpe or mangio-crepe for pitch detection — rmvpe is the default, chosen for vocal clarity, while mangio-crepe produces smoother vocals

This matters because a cover is a multi-step AI workflow, not a single model call: separate the stems, extract features, detect pitch, convert the voice, then remix. AICoverGen can run that entire chain in under a few minutes depending on your GPU.

The model ecosystem around RVC is just as important as the models themselves. Pre-trained RVC voice models are distributed through the AI Hub Discord community, meaning creators share and swap voice models across tools rather than relying on any single vendor. That community-driven distribution is a big reason RVC became the standard while commercial platforms like Jammable — which claims 60,000+ pre-made AI voices — keep their underlying architectures undisclosed.

For anyone orchestrating these multi-model pipelines, tool quality varies widely, which is exactly the problem SkillzBench is building around: verifying and benchmarking the agent skills that chain together steps like separation, conversion, and pitch correction. In a workflow where one weak link degrades the final vocal, knowing which components actually perform matters.

The takeaway: when someone asks what AI is used for music covers, the honest answer is a stack — RVC v2 for the voice, MDXNET for separation, HuBERT for features, and rmvpe or mangio-crepe for pitch — not one magic model.

Commercial Platforms: Jammable, Kits AI, ElevenLabs and Beyond

If you don't want to touch Python, GPUs, or Discord model-sharing, the commercial route gets you an AI cover in minutes — and the platforms have matured fast. The trade-off is cost and, in some cases, licensing clarity, but the quality gap with open-source tooling has largely closed.

Jammable (formerly Voicify AI) is the consumer front-runner. It offers more than 60,000 pre-made AI voices, claims content made with its voices has generated over 10 billion views across TikTok, YouTube, and Instagram, and trains custom voice models from a short vocal clip in minutes. Its most-used voices — Hatsune Miku (157,400 uses), Plankton (109,100), and a Michael Jackson re-upload (64,900) — show how cover culture straddles fictional characters and real artists. Note these are self-reported figures, so treat them as vendor claims rather than independently verified data.

Kits AI positions itself differently: studio-quality voice conversion built on ethically licensed artist models with revenue sharing for the original artists, starting at $10 per month. It bundles pitch correction, vocal isolation, stem separation, and AI mastering, making it a fuller production toolkit rather than a one-click novelty.

Other platforms worth knowing:

• ElevenLabs expanded from speech synthesis into music, voice cloning, and sound effects across 70+ languages.
• Soundverse Voice Swap replaces vocal identity while preserving timing, performance nuances, and expression, offering Community Voices and custom "DNA" voice identifiers.
• Controlla Voice and LALAL.AI cover voice conversion and stem separation for producers who want finer control.
• Open-source alternatives like Chatterbox (Resemble AI), Coqui XTTS, and OpenVoice round out the options for technical users.

One distinction matters: Suno and Udio are full-song generators, not cover tools. They create original music from text prompts — Suno's $250M Series C at a $2.45B valuation and $100M+ ARR reflect that market — whereas the platforms above convert existing vocal performances into new voices. If you want a cover, you want voice conversion; if you want an original track, you want generation.

The quality leap is real. Custom voice model training time has dropped from roughly 90 minutes three years ago to under five minutes today, with some platforms offering instant cloning from just 10–30 seconds of audio. Modern models capture timbre plus emotional inflection, breathing patterns, vibrato, and phrasing — which is partly why industry analysis finds 97% of listeners can't reliably identify AI-generated music.

Because these workflows chain multiple steps — separation, conversion, pitch correction, mixing — verifying which tools actually perform matters. That's the problem SkillzBench is building toward: independently benchmarking AI agent skills so creators know which pipeline components deliver and which underperform. And before publishing any cover, check licensing: models trained on copyrighted recordings can produce outputs that risk infringement claims, so consent-based platforms are the safer commercial path.

The Legal Shift: From Lawsuits to Licensed Voice Models

For anyone making AI covers in 2025, the technology is the easy part. The hard question is whether the voice model you're using belongs to you — or to someone who never said yes.

The legal ground has shifted fast. In 2024, the major labels sued Suno and Udio, with RIAA chief legal officer Ken Doroshow describing the cases as involving "unlicensed copying of sound recordings on a massive scale," as reported by the Hollywood Reporter. By late 2025, the same industry had pivoted from litigation to legitimization: Universal Music Group partnered with Udio in October 2025, and Warner Music Group settled with both Udio and Suno.

Cover-style outputs carry particular infringement risk. Legal analysis from IP firm Abou Naja notes that AI models trained on copyrighted music can produce outputs with "substantial similarity" to existing works — and purely AI-generated tracks often fall into the public domain, since human authorship remains a prerequisite for copyright protection.

The consent-based alternative is already here. Grimes's Elf.Tech platform, built with CreateSafe's TRINITI, offers a template: artists license their voice for a 50/50 split on master royalties, distribution costs $9.99 per year, and creators retain ownership of the underlying composition. Kits AI takes a similar approach, offering ethically licensed artist models with revenue-sharing, from $10 per month.

Two other trends are reshaping the risk picture:

• AI voice characters — original synthetic vocal identities that don't clone real people — are rising precisely because they sidestep licensing and consent issues entirely.
• Platform policies diverge sharply: Deezer labels and de-prioritizes AI music, while Spotify removed 75 million spam tracks in the 12 months through September 2025 without labeling AI content, and YouTube has no published AI music labeling policy.
• Detection infrastructure is layering up — training-data auditing (Fairly Trained, AI:OK), output attribution (ProRata.AI, Musical AI), and watermarking — but Water & Music's analysis cautions that deepfake detection will never reach full coverage and watermarking is easily circumvented.

Not everyone believes the settlement era serves artists. Irving Azoff, founder of the Music Artists Coalition, has warned that in label-AI partnerships, "artists end up on the sidelines with scraps."

For teams building AI audio workflows, the lesson is that provenance — where a voice model came from and who consented — now matters as much as output quality. It's the same principle SkillzBench applies to AI agent skills: verifying what's underneath a tool before you rely on it beats cleaning up after an infringement later.

Choosing Your Stack: A Practical Decision Framework

By now, the choice in AI cover tooling comes down to one question: do you want maximum control or minimum friction? The open-source and commercial paths both produce convincing covers — 97% of listeners can't reliably identify AI-generated music, per Deezer/Ipsos research across 8 countries — but they suit very different creators.

Pick the open-source RVC pipeline if budget is zero and control matters most. AICoverGen runs RVC v2 models locally, needs only Python, ffmpeg, and a GPU, and generates covers in a few minutes per track. You choose your pitch detector (rmvpe for clarity, mangio-crepe for smoothness), swap separation models, and own every output file. The tradeoff: you manage the legal risk yourself, and community-trained voice models raise real consent questions.

Pick a commercial platform if you need licensing safety and speed. Kits AI offers ethically licensed artist models with revenue-sharing from $10/month, while Jammable trains custom voices in minutes from a short clip. Given that AI models trained on copyrighted music can create infringement risk through "substantial similarity," per IP legal analysis, commercial platforms' licensed-model approach is the safer bet for anything you plan to release publicly.

A quick decision checklist:

• Choose RVC/AICoverGen when cost is the constraint, you have a GPU, and outputs stay personal or experimental.
• Choose Kits AI or Jammable when you need licensed voices, fast custom training, or professional release quality.
• Consider consent-based models like Grimes's Elf.Tech (50/50 royalty split) if you want commercial distribution with a clear legal template.
• Explore original synthetic voice characters to sidestep licensing and consent issues entirely.

Whichever stack you choose, remember that an AI cover is never one model — it's a pipeline: separate the vocals (MDXNET), convert the voice (RVC v2), detect pitch (rmvpe), and mix the result. Each step is a distinct skill an AI agent must orchestrate correctly, and a weak link anywhere degrades the final track.

That's the problem SkillzBench is building toward: when agent skills promise to automate multi-model workflows like these, someone needs to prove they actually work. SkillzBench is creating the benchmarking and verification layer — testing every skill in a sandbox, with and without it, and publishing the measured Lift — so you can trust the skills orchestrating your pipeline. Skills, proven — not just promised. Join the waitlist at skillzbench.com to see it take shape.