Sorinel Stems
Our in-house AI model for audio stem separation. Splits any track into vocals, drums, bass and other — running 100% on-device on Apple Silicon.
PLAY
1 Model
~5.5x
44.1kHz
4 Stems
Meet SorinelStems
A proprietary neural network family for music stem separation.
SDR — Distortion
SIR — Interference
SAR — Artifacts
SDR
SIR
SAR
Separation Quality (dB)
Per-stem reconstruction quality measured on an internal held-out validation set. SDR reflects overall separation; SIR measures clarity between stems; SAR measures reconstruction smoothness. Higher is better. Results naturally vary with genre, mix complexity, and source audio quality.
Tests performed on 4 AI-generated songs from 2026.
100% on-device
Apple Silicon
~5.5× realtime on MPS
Integrated in: