MVSEP is a web platform that runs more than a hundred AI source-separation models, letting you split a song into stems with whichever model does the cleanest job on that particular track.
Most separators give you one algorithm and one result. MVSEP instead hosts the leading open and proprietary separation models ā including the BS RoFormer and MelBand RoFormer families, MDX23C, HT Demucs and SCNet ā and lets you choose between them or compare outputs. Beyond vocal and instrumental splits, dedicated models isolate drums, bass, guitar, piano and even niche instruments, generate karaoke stems (lead vocal, backing vocal, instrumental), remove reverb, or enhance and upscale audio. Processing happens in the cloud with exports to MP3, WAV and FLAC, there are mobile apps and a full API, and the service’s privacy policy states that uploaded audio is never used to train models.
The free tier allows 50 separations per day on files up to 10 minutes and 100 MB. Premium credits unlock files up to 100 minutes, unlimited daily separations, batch processing, lossless exports and full API access.
MVSEP also maintains public quality benchmarks comparing its models on synthetic and multi-song datasets, which is why it is popular with researchers as well as musicians. Premium tiers add ensemble processing that combines several models for tracks where a single model leaves artifacts ā and the plain, technical interface assumes you know what a stem is. Power users would not have it any other way.
Producers, remixers, DJs, karaoke creators and audio researchers who work with difficult material and want model choice and quality control instead of a single black-box result.