Five years ago, isolating a clean vocal from a finished mix was a parlour trick. Today it ships inside Logic Pro and powers entire stem libraries at major labels. The path from research curiosity to commodity tool was short, and the architectures got smarter in a specific direction worth understanding.
From U-Nets to Transformers
The first widely usable model was Spleeter, released by Deezer in late 2019. Architecturally it is a U-Net convolutional network with twelve layers, six in the encoder and six in the decoder, predicting a soft mask on the STFT magnitude spectrogram. Multiply the mask by the mixture spectrogram, run an inverse STFT, and you have your stems. Spleeter was fast, up to 100 times real-time on a GPU, and it still works. But Deezer essentially stopped active development around 2022, and modern open models now beat it by roughly 3 dB SDR. That is not a small gap.
The big leap came with Demucs. The original Demucs from Meta AI went the opposite direction to Spleeter and operated directly on the waveform, on the theory that staying in the time domain preserves phase information that spectrogram-based models throw away. Demucs v3 fused both approaches into a Hybrid model, and Demucs v4, also called HTDemucs (Rouard, Massa, Défossez, arXiv 2211.08553), took this further. Two parallel U-Nets, one waveform, one spectrogram. After the fifth layer their latent representations are summed, and the innermost block is a Transformer that does self-attention within each domain and cross-attention between them. On MUSDB-HQ, HTDemucs reaches about 9.00 dB SDR without extra training data, and 9.20 dB with additional data and per-source fine-tuning.
BS-RoFormer from ByteDance SAMI (Lu et al., arXiv 2309.02612) is currently the model to beat. It returns to spectrogram-only processing but factorises the problem differently. The complex spectrogram is split into K subbands, each projected into a fixed-dimension representation. Two Transformer blocks alternate. An inner-band block models the time sequence within a single band. An inter-band block models the band sequence at each time frame. Rotary Position Embedding (RoPE) gives the attention mechanism relative positional information. A smaller BS-RoFormer hits around 9.80 dB SDR on MUSDB18-HQ without extra training data, already state of the art. With MUSDB18-HQ plus 500 additional songs the model took first place on the SDX23 Music Demixing track, beating the 2021 MDX winner by more than 1.6 dB.
Mel-Band RoFormer (arXiv 2310.01809), also from ByteDance and only a month later, swapped the linear band-split for mel-scaled overlapping subbands. The change is small in code and biologically motivated. It outperforms BS-RoFormer on vocals, drums and the other stem.
What SDR Numbers Actually Mean
Signal-to-Distortion Ratio is the standard metric. Compute it per second of audio, then average. MUSDB18 is the benchmark, 150 songs split 100 for training and 50 for testing, with stems for vocals, drums, bass and other. MUSDB18-HQ is the uncompressed version of the same set.
The honest variant is SI-SDR, which is scale-invariant. Plain SDR punishes scale and time-shift errors that human listeners do not actually hear, and it rewards conservative outputs that leave parts of the signal alone. A 2025 paper, the Musical Source Separation Bake-Off (arXiv 2507.06917), compared SDR rankings with formal listening tests and found that the two can disagree substantially. So when product reviews and leaderboard tables put the same models in different orders, both can be right. They are measuring different things.
A rough rule of thumb, with caveats: every 3 dB of SDR is perceptually noticeable on the cleanest reference material, but the relationship is not linear and not consistent across stems. Vocals are easier to separate cleanly than the other bucket, which holds whatever the model could not file elsewhere.
What Ships in Products
Logic Pro 11 (May 2024) shipped Apple's own Stem Splitter, and the 11.2 update extended it to six stems: vocals, drums, bass, guitar, piano and other. Apple has not published the architecture. The popular claim that it is a HTDemucs variant is not verified, so it is best treated as a black box that happens to work well. In a 2025 MusicRadar shootout of eleven stem separation tools, Logic's offline Stem Splitter ranked at or near the top for vocal isolation quality.
iZotope RX 12 Music Rebalance lives in the audio repair workflow rather than music production. AudioShake powers LANDR Stems and is what labels and publishers tend to use for sync work and licensed-music re-releases. Moises has the friendliest UX with key, tempo and click-track features, though side-by-side tests still hear characteristic warbling artefacts on its vocal stem.
For free and open, the practical entry points are UVR (Ultimate Vocal Remover) and StemRoller. Both bundle the open Demucs and RoFormer-family models. The ceiling is higher than any cloud service, but you need a GPU with roughly 4 to 8 GB of VRAM and you should expect 20 to 60 seconds of processing per three-minute song.
Where It Still Bleeds
Every current model leaks. The two persistent failure modes are cymbal bleed into the other stem and reverb tails following the vocal. The second one matters in remix work. If you wanted the dry vocal gone but the room intact, current models will not give you that cleanly. They tend to follow the vocal energy and take the reverb with it.
Phase coherence between stems is generally good for modern networks. You can sum the separated stems back to the original mix without significant phase artefacts, which suggests the networks are not just hiding signal in arbitrary phase combinations. That matters if you plan to process stems separately and recombine them.
A Quiet Convergence
Two architectural ideas keep appearing. The first is dual-domain processing, running waveform and spectrogram paths in parallel and fusing them. The second is frequency-band factorisation, processing narrow bands independently and then letting the network exchange information across them.
Both have a real biological counterpart. The cochlea performs a place-code analysis of frequency, organised tonotopically, which is the band-factorisation principle. The auditory nerve also phase-locks to the time-domain waveform up to roughly 4 kHz, which is closer to the waveform-domain branch. Modern stem separation networks did not set out to imitate the ear, but they have converged on a similar split-and-combine strategy, more or less by accident. That is probably the most interesting thing about the current state of the art. Not that the models are bigger, which they are, but that they finally stopped picking between two representations and started using both.