Introducing Multi Stem HQ

Multi Stem HQ is built upon the advanced Band-Split architecture. Unlike traditional models that process the full spectrogram as a single image, this technology divides the audio into multiple frequency bands.

This allows the model to learn distinct features for each frequency range—capturing the deep resonance of the bass independently from the intricate harmonics of vocals. The result is a dramatic reduction in spectral leakage and artifacts.

Furthermore, the architecture integrates Rotary Position Embeddings (RoPE). This mathematical innovation allows the transformer to understand the relative position of audio features across time more effectively than standard absolute position embeddings. It enables the model to maintain coherence over longer musical phrases, ensuring that transient sounds like drum hits are sharp and sustained notes are smooth.