AES Europe 2026

This paper presents a multitrack dataset designed to
support music production research; education, including
machine learning techniques such as automatic mixing;
source separation. The dataset comprises a cohesive 10-song
indie album (indie rock/folk), with separate stems for
individual instruments, such that each song has between 13
; 35 individual tracks (stems). For each song, three
versions of each stem are provided: the raw unprocessed
stems, a dry mixed version (processed but without
reverberation or delay effects),; a full mixed version.
Additionally, each song includes two final master formats:
stereo; immersive 7.1.4. This album-format dataset
enables studies of mix consistency across a thematically
aligned collection of songs, as well as stereo upmixing to
immersive formats,; contains far more stems per song
than traditional four-stem datasets. To illustrate an
example usage of the dataset, the MEGAMI automatic mixing
model is used to produce a mix for two songs. The results
are analysed in comparison to the raw (unmixed); human
mixed versions. The dataset is made open-access; free to
download.