AI Stem Separation

Stem separation (source separation) uses AI models to decompose a mixed audio signal into individual components — vocals, drums, bass, and other instruments. This is one of the most practically useful AI audio technologies for music producers.

What Is Stem Separation?

Given a mixed audio signal $x[n]$, the goal is to recover the source signals:

$$x[n] = s_{\text{vocals}}[n] + s_{\text{drums}}[n] + s_{\text{bass}}[n] + s_{\text{other}}[n]$$

This is an underdetermined problem — there are infinitely many valid decompositions. AI models learn statistical priors over what each source "should" sound like.

How AI Separation Works

Spectrogram Masking

The dominant approach estimates a mask for each source in the time-frequency domain:

$$\hat{S}_k(t, f) = M_k(t, f) \cdot X(t, f)$$

where $M_k$ is the estimated mask for source $k$ and $X$ is the mixture spectrogram.

The masks should sum to 1 at each time-frequency bin (or approximately so):

$$\sum_{k} M_k(t, f) \approx 1$$

Mask Types

Mask Type	Range	Properties
Binary mask	1	Hard assignment, artifacts
Ratio mask	[0, 1]	Soft, smoother transitions
Complex mask	$\mathbb{C}$	Handles phase, best quality

Waveform-Domain Approaches

Some models operate directly on waveform:

$$\hat{s}_k[n] = f_\theta^{(k)}(x[n])$$

This avoids phase estimation issues but requires the model to learn both magnitude and phase separation.

Key Models

Demucs (Meta)

Hybrid Transformer Demucs (HTDemucs) is the current state-of-the-art open-source separator:

• Architecture: Hybrid of waveform U-Net + spectrogram U-Net with transformer layers
• Sources: vocals, drums, bass, other (4-stem)
• Quality: Excellent, especially for vocals
• Open source: Available in Meta's demucs repository

The hybrid approach processes both waveform and spectrogram in parallel, combining their outputs:

$$\hat{s}_k = \alpha \cdot \text{WaveformDecoder}_k + (1-\alpha) \cdot \text{SpectrogramDecoder}_k$$

Open-Unmix

• Bidirectional LSTM on spectrogram
• Open source, well-documented
• Good baseline, outperformed by Demucs

Band-Split RNN (BSRNN)

• Splits spectrogram into frequency bands
• Applies band-specific RNNs
• Won the 2023 Music Demixing Challenge
• Excellent performance, especially for vocals

Music Source Separation Transformer (MSSM)

• Pure transformer operating on spectrogram patches
• Attention across both time and frequency
• Competitive with Demucs

Quality Metrics

Signal-to-Distortion Ratio (SDR)

The standard metric for source separation:

$$\text{SDR} = 10 \log_{10} \frac{\|s_{\text{target}}\|^2}{\|s_{\text{target}} - \hat{s}\|^2}$$

Higher SDR = better separation. Measured in dB.

Scale-Invariant SDR (SI-SDR)

Removes amplitude scaling as a factor:

$$s_{\text{proj}} = \frac{\langle\hat{s}, s\rangle}{\|s\|^2} s$$ $$\text{SI-SDR} = 10 \log_{10} \frac{\|s_{\text{proj}}\|^2}{\|\hat{s} - s_{\text{proj}}\|^2}$$

Typical Performance (SDR in dB)

Source	Demucs v4	BSRNN	Baseline
Vocals	~8.5 dB	~9.0 dB	~5.5 dB
Drums	~8.0 dB	~7.5 dB	~5.0 dB
Bass	~7.0 dB	~7.0 dB	~4.5 dB
Other	~5.5 dB	~5.5 dB	~3.0 dB

"Other" (containing everything else) is hardest because it's the most diverse category.

Practical Applications for Producers

1. Remixing AI Outputs

Separate your AI-generated track into stems, then:

• Replace weak drums with better samples
• Rebalance the mix (louder vocals, quieter bass, etc.)
• Apply different effects to each stem
• Remove unwanted elements

2. Creating Mashups

Combine elements from different AI generations:

• Vocals from generation A
• Drums from generation B
• Bass and harmony from generation C

3. Fixing Mix Problems

• Extract and fix problematic elements
• Apply targeted EQ or compression per stem
• Remove artifacts from specific stems

4. Sampling and Sound Design

• Extract interesting sounds from AI generations
• Build sample libraries from separated elements
• Use isolated elements as layers in traditional productions

5. A Cappella and Instrumental Creation

• Extract vocals for a cappella versions
• Create instrumental versions by removing vocals
• Isolate specific instruments for analysis or reuse

Workflow: AI Generation + Stem Separation

Step 1: Generate with AI music platform
    │
    ▼
Step 2: Select best generation
    │
    ▼
Step 3: Separate into stems (Demucs)
    │
    ├──▶ Vocals
    ├──▶ Drums
    ├──▶ Bass
    └──▶ Other (synths, guitars, etc.)
    │
    ▼
Step 4: Process stems individually
    │
    ├──▶ EQ, compression per stem
    ├──▶ Replace weak elements
    └──▶ Add your own layers
    │
    ▼
Step 5: Remix and master

Limitations and Artifacts

Bleeding

Imperfect separation leaves traces of other sources in each stem:

• Vocal bleed in drum stem
• Drum transients in vocal stem

Mitigation: use higher-quality models, accept some bleed, or manually edit.

Phase Artifacts

Separation can introduce phase distortion:

• Hollow or phasey sound when recombining stems
• Worse in spectrogram-domain methods than waveform methods

Quality Degradation

Each separated stem is lower quality than the original mix:

• Some information is inevitably lost
• Artifacts increase with lower SDR
• "Other" category typically has most artifacts

Processing Tips

• Don't over-process separated stems — artifacts become more audible
• Use gentle EQ to reduce bleed rather than aggressive gating
• Layer with fresh sounds to mask separation artifacts
• Check mono compatibility — separation can create stereo issues