Separation quality built for real listening conditions.
Demixr Pro is tuned for real music: heavy reverb, dense arrangements, low-bitrate compression, and live ambience. This lab will collect separation results, listening samples, and model iteration notes.
Demixr Pro estimates time-frequency ownership, not just noise reduction.
The model decomposes music into a high-resolution spectrum, reorganizes it into perceptual energy bands, then estimates whether each region belongs to vocal, accompaniment, or overlap. A multi-scale consistency pass keeps vocal edges natural and reduces bleed.
High-resolution spectral front-end
44.1kHz stereo analysis with a 2048-point spectral window and fine hop size captures sibilance, transients, and low-end energy.
Adaptive band partitioning
Narrow low bands and wider high bands let the model read vocal fundamentals, harmonics, and accompaniment texture together.
Dual-axis time-frequency attention
A 12-layer separation core follows syllable continuity over time and resolves instrument masking across frequency.
Multi-scale consistency
Multiple spectral scales reduce watery artifacts, metallic edges, and smeared transients for listening and post-production.
A/B listening tests from real samples.
The samples below come from the local mp3 folder and have been separated with Demixr Pro. Each card lets you compare original, vocal, and instrumental outputs.