How VocRemover Works

This page explains what happens to a file after you upload it: which models do the separation, what we do to your audio and video along the way, and where the results are not perfect. We keep it up to date whenever the pipeline changes.

Two tools, two models

Separating a song and separating a video soundtrack are different problems, so VocRemover uses a different model for each.

Songs: Demucs (htdemucs_ft)

The voice remover uses htdemucs_ft, the fine-tuned Hybrid Transformer variant of Demucs, an open-source music source separation model published by researchers at Meta AI. Demucs analyses the waveform and the spectrogram of a song together, which lets it separate a voice from instruments that occupy the same frequency range.

We run it in two-stem mode, so every song comes back as:

• Vocals — lead and backing vocals together.
• Instrumental — everything else in the mix.

Video soundtracks: DBS-TIGER

The video music remover uses DBS-TIGER, a separation model trained on the Divide and Remaster (DnR) dataset. DnR was built for film and TV audio, where the job is not “voice vs. band” but dialogue vs. music vs. sound effects. Every video soundtrack comes back as three layers:

• Dialogue — speech.
• Background music — scored or licensed music, including sung lyrics.
• Other sounds — effects, ambience, applause, room tone.

What happens to your file

1. Upload. You upload from your browser over HTTPS. Audio files can be up to 80 MB (MP3, WAV, FLAC, M4A, AAC, OGG); videos up to 90 MB (MP4, MOV, WEBM, M4V). We check that the file’s contents actually match its extension before processing.
2. Private storage. The file is stored in a private Cloudflare R2 bucket. It is not publicly listed; processing workers and your browser access it through short-lived signed links.
3. Separation on a GPU. Songs are processed through Replicate; videos are processed on our own GPU worker running on RunPod. For video, the soundtrack is first converted to 44.1 kHz audio before separation.
4. Rebuilding the video. The layers you chose to keep are mixed together and placed back into a copy of your video. The picture stream is copied without re-encoding, so its resolution and quality are unchanged. The new audio track is AAC at 192 kbps.
5. Results. You can play every track in the workspace before downloading, and your past jobs are listed under My files in your account.

Known limitations

We would rather you know these before you upload:

• Reverb and delay. Heavy effects on a vocal can leave a faint trace of the voice in the instrumental.
• Backing vocals. Harmonies and backing vocals are part of the vocal stem; they can’t yet be kept separately from the lead.
• Voice-like instruments. Lead synths, saxophones, and vocoders can partly bleed into the vocal stem.
• Low-quality sources. Low-bitrate MP3s and audio ripped from compressed video separate less cleanly, because compression artifacts are split along with the music.
• Mono video audio. The video model works on a mono version of the soundtrack, so the finished video’s audio is centered rather than stereo.
• Singing in video. Because the video model treats singing as music, sung lyrics usually go to the music layer, not the dialogue layer.

Your rights and responsibilities

Separating a recording does not change who owns it. Only upload files you have the right to process, and get permission before publishing or monetizing work based on someone else’s recording. Removing a music track from a video is a cleanup tool, not a way around a licence.

Questions or a bad result?

If a file didn’t separate the way you expected, email hello@vocremover.com with the job name from My files. Hard examples help us decide what to improve next.