How much quality does a faster vocal remover cost?
Some vocal removers finish a song in seconds; others take minutes. We measured six open separation models on 50 test songs to see what the extra time buys, what the difference sounds like and what else decides how clean your result will be.
The short answer
The fastest model we measured worked more than 200 times faster than the slowest, but its vocals were about 5 dB less accurate than the best result. That is a difference most people hear at once, as drums and guitars bleeding into the voice. Between the mid-sized models and the largest one, the gain shrank to 1–2 dB, for two to eleven times the processing time.
In practice, a very fast remover is fine for a rough idea or a practice track. For a remix, a cover or a karaoke track you will share, the extra processing time is usually worth it.
What we tested
We used MUSDB18, a public research collection of songs released together with their separate vocal, drum, bass and other tracks, so a separated vocal can be compared with the real one. We ran each model on 7-second excerpts from the 50 songs in its test set, measured the signal-to-distortion ratio (SDR) used in the Music Demixing Challenge, and report the median over the 50 songs.
Speed is the number of seconds of processing for each second of audio on a recent laptop processor (Apple M4, four threads). A value of 0.5 means a three-minute song takes about a minute and a half.
Results
| Model | Vocal SDR | Instrumental SDR | Time per second of audio |
|---|---|---|---|
| Spleeter, 2 stems | 6.5 dB | 11.3 dB | 0.02 s |
| SCNet small | 9.8 dB | 14.1 dB | 0.37 s |
| DTTNet | 10.4 dB | 14.7 dB | 0.87 s |
| SCNet large | 10.7 dB | 15.1 dB | 1.93 s |
| SCNet XL IHF | 11.2 dB | 15.3 dB | 4.73 s |
| Large transformer model | 11.7 dB | 15.8 dB | 4.07 s |
- Spleeter is very fast, but its output stops at 11 kHz, so vocals sound dull.
- SCNet small, SCNet large and SCNet XL IHF also separate drums, bass and other instruments; SCNet XL IHF gave the best drums and other instruments in this group.
- DTTNet separates vocals only, and its vocal output stops near 14.7 kHz.
- The large transformer model separates vocals and instrumental only, and gave the cleanest vocals in this test.
Every model scored higher on the instrumental than on the vocal. The instrumental carries more of the song’s energy, so the same amount of leftover sound costs it less. For a karaoke track or a beat, watch the instrumental score; for a remix or an a cappella, watch the vocal score.
Size alone did not decide quality. The smallest vocal model in the test came within 1.4 dB of the best vocals, and the slowest model was not the most accurate. How a model is designed and trained matters as much as how big it is.
What the differences sound like
| Vocal SDR | What you typically hear |
|---|---|
| Around 6–7 dB | Clear traces of drums and guitars in the vocal, a watery or phasey sound, and a dull top end when the output stops early. |
| Around 10 dB | Usable for most remixes. Faint hi-hats, cymbals and reverb tails can remain in quiet passages. |
| Around 11–12 dB | Leftovers mostly limited to breaths, sharp consonants and effects that belong to both parts. |
These descriptions are a rough guide. The same score can sound different on two songs, because the score averages every kind of error into one number.
What else decides how clean your result is
- Source quality. A lossless or high-quality file separates more cleanly than a low-quality download, because compression blurs the details a model uses to tell voice from music.
- Reverb and delay. Echoes of the voice can be treated as part of the vocal or part of the music, so a very wet vocal can leave a ghost behind in the instrumental.
- Backing vocals and ad-libs. They count as vocals, so a vocal remover usually takes harmonies out of a karaoke track too.
- Instruments in the vocal range. Saxophone, lead synths and vocal samples can be partly treated as voice.
- Older and mono recordings. When the voice and the band share one channel and a narrow range of frequencies, every model has less to work with.
- The kind of mix. These scores come from studio mixes. Live recordings, crowd noise and very dense masters can score lower.
How to get a cleaner result
- Start from the best file you have. A WAV or FLAC master beats a streamed or re-encoded copy.
- Listen to both outputs in the loudest chorus, where bleed is most obvious, and in a quiet verse, where artefacts are easiest to hear.
- Check a karaoke track on headphones. Leftover vocals are easier to hear there than on speakers.
- If you only need part of a song, trim it first. Shorter files finish sooner and are easier to check.
To trim first, use Audio Cutter. If you need drums or bass as well as the vocal, use Stem Splitter instead of separating the song twice.
Vocal Remover and Stem Splitter
Vocal Remover separates a song into a vocal track and an instrumental, for karaoke, practice or an a cappella. Stem Splitter separates vocals, drums, bass and other instruments for remixing and study.
Building a remix from the instrumental? Read why a BPM finder shows half or double the tempo and how accurate automatic song key detection is before you match it to another track.
References
MUSDB18: Rafii et al., the MUSDB18 corpus for music separation, 2017.
SDR: the whole-song signal-to-distortion ratio from Mitsufuji et al., “Music Demixing Challenge 2021”, Frontiers in Signal Processing, 2022.
Spleeter: Hennequin et al., “Spleeter: a fast and efficient music source separation tool with pre-trained models”, Journal of Open Source Software, 2020.
SCNet: Tong et al., “SCNet: Sparse Compression Network for Music Source Separation”, ICASSP 2024, and its published code. SCNet XL IHF weights: the Music Source Separation Training project.
DTTNet: Chen et al., “Music Source Separation Based on a Lightweight Deep Learning Framework (DTTNet)”, ICASSP 2024.