AI Vocal Remover

AI vocal removal uses a deep learning model (Mel-Band RoFormer) that analyzes the spectrogram of your song across time and mel-scaled frequency bands. It estimates the vocal track, and the instrumental is the original mix with that vocal track removed, producing clean karaoke-style output.

The vocal model measured 11.3 dB vocal SDR and 16.0 dB instrumental SDR on 50 MUSDB18 test excerpts, about 3 dB better on both than the engine we used before. The instrumental output is typically clean enough for karaoke or remixing. Some bleed-through may occur on heavily processed or compressed tracks.

Yes! You get both the isolated vocals and the instrumental track. This is useful for creating acapella versions, sampling vocals for remixes, or transcribing lyrics from complex mixes.

We support MP3, WAV, FLAC, OGG, M4A, and WEBM files up to 50MB. For best results, use the highest quality source file available (WAV or FLAC preferred over compressed MP3).

Yes, creating karaoke tracks is one of the most popular uses. The instrumental output is clean enough for karaoke performances, sing-along videos, and background music for events.

Upload the song and the tool automatically produces both the isolated vocal track (acapella) and the instrumental. The vocal extraction works best on well-mixed tracks with clear vocals. Use WAV or FLAC source files for the cleanest results.

Batch processing is available through our API, letting you submit multiple tracks for vocal removal in an automated workflow. The web interface processes one song at a time with instant preview and download.

Our tool uses a Mel-Band RoFormer model for vocal separation. On 50 MUSDB18 test excerpts it measured 11.3 dB vocal SDR and 16.0 dB instrumental SDR. Transformer-based separation like this leaves fewer artifacts than older spectral-based methods.

Yes, but results vary with live audio. Studio recordings give the best separation quality. Live recordings with audience noise, reverb, and overlapping instruments may have more bleed-through, though the model still produces usable results.

Vocal removal uses 2,000 characters per track processed. Free accounts receive 15,000 characters on signup. The tool is available on all paid plans with generous character allowances for regular use.

Audio separated with Spleeter on a paid plan can be used commercially, and upgrading also covers audio you separated earlier on the free tier; free-tier output is for personal use. The Demucs model, the default for best quality, has weights released for scientific use only, so treat Demucs output as personal and non-commercial use only on every plan. In every case you must have the rights to the original song: using copyrighted music without permission may violate copyright laws regardless of how the audio is processed.

Processing time depends on the track length. A typical 3-5 minute song takes 15-45 seconds to process. Longer tracks or higher-quality separation modes may take up to a minute. Results play back instantly in your browser.