ACMID: Automatic 7-Stem Music Dataset

• ACMID is a dataset and end-to-end curation pipeline for music source separation that provides high-purity audio for seven distinct instrument stems.
• It employs multilingual YouTube crawling, safe normalization, and Dasheng-based classifiers to filter 4,643+ hours of raw audio down to 15.9% clean segments.
• ACMID enhances model performance with significant SDR gains and offers a fine-grained taxonomy, enabling improved data augmentation and reproducible research in MSS.

ACMID is a dataset and end-to-end curation pipeline for music source separation (MSS) that combines web-scale collection with automatic cleaning to obtain high-purity single-instrument audio for seven target stems: Piano, Drums, Bass, Acoustic Guitar, Electric Guitar, Strings, and Wind-Brass. It was introduced in "ACMID: Automatic Curation of Musical Instrument Dataset for 7-Stem Music Source Separation" (Yu et al., 9 Oct 2025) to address a central limitation of supervised MSS, namely that existing performance is constrained by the quantity and quality of training data, while naïve web crawling suffers from platform-level track labeling errors and consequent metadata mismatches. ACMID therefore distinguishes between a large raw corpus, ACMID-Uncleaned, and a filtered subset, ACMID-Cleaned, produced by a classifier-driven segment selection procedure.

1. Motivation and problem setting

Most current MSS methods rely on supervised learning, and their efficacy is therefore bounded by the scale and fidelity of available training corpora (Yu et al., 9 Oct 2025). Web crawling can provide abundant audio, but platform metadata is not sufficiently reliable for direct supervision because track labels and acoustic content may not align. In the ACMID framework, this problem is framed as the acquisition of accurate "audio-label" pairs from noisy online sources.

A second motivation is representational granularity. Conventional 4-stem settings partition mixtures into Vocal, Bass, Drums, and Others. In that formulation, acoustic guitars, electric guitars, strings, and wind/brass are merged into a single residual category. The ACMID design argues that this compression prevents models from learning instrument-specific timbres, and instead curates seven distinct stems so that MSS systems can specialize at a finer level of musical structure (Yu et al., 9 Oct 2025).

2. Automatic curation pipeline

The ACMID pipeline begins with web crawling from YouTube using the query template "<instrument> solo", localized into nine languages to improve coverage (Yu et al., 9 Oct 2025). A metadata filter keeps only videos whose title or description matches the multilingual “instrument + solo” pattern. The downloaded audio is converted to 2-channel (stereo) 48 kHz, producing ACMID-Uncleaned, which totals 4 643.51 hours across seven instrument categories.

Cleaning is performed by seven binary classifiers, one per stem, each trained to distinguish “clean” from “impure” 3 s excerpts. The classifier input is a 3 s mono waveform at 16 kHz, with safe normalization

$$x \leftarrow \frac{x}{\max\bigl(|x|\bigr)+\varepsilon},\quad \varepsilon=10^{-9}.$$

The backbone is a frozen pre-trained audio encoder, Dasheng, described as a self-supervised masked-autoencoder audio model. Its sequence of embeddings is averaged over time to obtain a $D$-dimensional feature vector, which is passed to a three-layer fully connected head with Linear→ReLU transformations of sizes

$D \to \tfrac{D}{2} \to \tfrac{D}{4},$

followed by a final linear layer that outputs a scalar logit. The posterior and decision rule are

$$p = \sigma(\text{logit}),\quad \hat y = \begin{cases} 1 & p\ge\tau,\ 0 & p<\tau, \end{cases}$$

with $\tau=0.5$ during training and $\tau=0.995$ at inference. Optimization uses standard binary cross-entropy,

$$\mathcal{L}(\hat p,y) = -\bigl[y\log \hat p + (1-y)\log(1-\hat p)\bigr].$$

The classifier supervision is constructed from isolated-stem crops and synthetic negatives. Positive samples are 3 s crops of the isolated instrument from MedleyDB, MoisesDB, Bach10, ARME-Strings, SynthTab, and MAESTRO. Negative samples mix $k\in[1,5]$ other instruments and noise. Augmentations include random reverb with $\mathrm{RT60}\in[0.3,1.4]\,\mathrm{s}$, dynamic compression, EQ boosts, and LUFS normalization (Yu et al., 9 Oct 2025).

At inference time, raw audio is segmented into contiguous 3 s chunks, each chunk is scored by the corresponding instrument classifier, and a segment is marked “clean” if $p\ge 0.995$. All clean segments for a given instrument are concatenated into the final per-instrument track. This procedure retains 737.35 hours of high-purity single-instrument audio from the original 4 643.51 hours, or approximately 15.9% of the raw corpus (Yu et al., 9 Oct 2025).

3. Stem taxonomy and corpus composition

ACMID adopts a seven-stem taxonomy comprising Piano, Drums, Bass, Acoustic Guitar, Electric Guitar, Strings, and Wind-Brass (Yu et al., 9 Oct 2025). The grouped classes are explicitly defined. Strings consists of Cello, Viola, Violin, and Double Bass. Wind-Brass consists of Trombone, Trumpet, Tuba, Euphonium, French Horn, English Horn, Bassoon, Clarinet, Contra Bassoon, Flute, Oboe, Piccolo, and Saxophone.

The paper reports only aggregate totals as formal corpus statistics, but Figure 1 provides an approximate per-stem duration breakdown. Those approximate values are as follows.

Instrument	Uncleaned	Cleaned
Piano	1 000 h	150 h
Drums	800 h	140 h
Bass	720 h	120 h
Acoustic Guitar	400 h	70 h
Electric Guitar	500 h	80 h
Strings	650 h	90 h
Wind-Brass	573 h	87 h

The total duration is 4 643.51 h for ACMID-Uncleaned and 737.35 h for ACMID-Cleaned (Yu et al., 9 Oct 2025). The cleaning process is described as uniformly removing impure or mixed-instrument segments, reducing each stem to approximately 15–25% of its uncleaned duration. Post-cleaning, all seven stems retain at least 70 hours of data. This suggests that the filtering threshold is tuned toward purity rather than raw volume, while still preserving enough material to train high-granularity separation systems.

4. From 4-stem to 7-stem music source separation

The ACMID paper situates its contribution against the conventional 4-stem formulation of MSS, in which all non-vocal, non-bass, and non-drum material is assigned to “Others” (Yu et al., 9 Oct 2025). That design choice collapses heterogeneous timbral families into a single residual class. In the ACMID framing, this makes it difficult for a model to learn distinctions such as acoustic versus electric guitar or strings versus wind/brass.

By curating seven stems instead of four, ACMID is intended to support higher-granularity source separation. The paper explicitly identifies the ability to specialize on categories such as “Wind-Brass” versus “Strings,” and to disentangle acoustic and electric guitars, as a direct consequence of the new taxonomy. It further links this granularity to downstream usefulness in remixing and intelligent mixing assistants (Yu et al., 9 Oct 2025).

A plausible implication is that ACMID is not merely a larger training set, but a restructuring of the target space for MSS. The gain therefore lies both in dataset scale and in a more discriminative label ontology.

5. Experimental validation with SCNet

Experimental evaluation is conducted with SCNet as the state-of-the-art MSS model (Yu et al., 9 Oct 2025). The separation model takes 2-channel, 10 s audio at 48 kHz as input. Its front end uses an STFT with a 4 096-point FFT, 1 024 hop, and 4 096 window. The filterbank is a hierarchical band-split with three subbands at sampling ratios 0.230/0.370/0.400 and per-band depthwise convolutions of depths 3/2/1. The encoder/decoder has 8 down/up-sampling layers with factor 4. Training uses Adam with learning rate $D$0, batch size 8, mixed precision, maximum 1 000 epochs, and learning-rate reduction factor 0.95 with patience 2. The train/test splits are the same validation/test sets of MedleyDB and MoisesDB used in classifier training.

Performance is measured with signal-to-distortion ratio (SDR), following Vincent et al. (2006):

$D$1

$D$2

Two quantitative results are emphasized. First, training on ACMID-Cleaned rather than ACMID-Uncleaned increases average SDR from 2.24 dB to 4.63 dB, a gain of 2.39 dB (Yu et al., 9 Oct 2025). This is presented as evidence that the automatic cleaning procedure materially improves supervisory quality. Second, augmenting a MoisesDB + MedleyDB training set with ACMID-Cleaned increases average SDR from 4.89 dB to 6.05 dB, a gain of 1.16 dB. This is presented as evidence that ACMID-Cleaned contributes useful additional training signal beyond existing datasets.

6. Research uses, reproducibility, and interpretive boundaries

The paper recommends ACMID for fine-grained 7-stem MSS research and model benchmarking, for data augmentation or pretraining in architectures targeting instrument-level disentanglement, for domain adaptation or self-supervised refinement using automatically cleaned pseudo-labels, and for curriculum learning in which training is warm-started on ACMID-Cleaned before noisier data is introduced (Yu et al., 9 Oct 2025).

Reproducibility is a central operational feature. The project releases all web-crawling scripts, binary classifier training and inference code, pre-trained Dasheng-based classifiers, SCNet separation training scripts, and pre-trained weights at the ACMID repository. The pipeline from crawling to cleaning to separation is therefore provided end to end (Yu et al., 9 Oct 2025).

Several interpretive boundaries are important. ACMID is generated through web crawling and automatic cleaning rather than through direct manual annotation of every retained segment. The cleaned subset is therefore a filtered selection from a much larger noisy corpus, not an exhaustive representation of all crawled material. The retention of only approximately 15.9% of the raw audio further indicates that purity is enforced by a stringent decision rule at inference, specifically the threshold $D$3. This suggests that ACMID-Cleaned should be understood as a high-precision subset designed to optimize label fidelity for MSS training rather than as a maximal-coverage archive of online instrumental audio.