WCMED: Western Classical Music Dataset

• WCMED is a curated corpus of 400 Western classical music excerpts designed to test the transferability of soundscape models for music emotion recognition.
• The dataset features short 8–20 second WAV excerpts with pairwise ranking annotations for valence and arousal, ensuring timbral consistency and methodological rigor.
• WCMED serves as a diagnostic tool for distribution shift, revealing that while arousal is effectively captured by timbral cues, valence prediction remains challenging for soundscape-based models.

Searching arXiv for papers on “WCMED” and its direct research context. Searching arXiv for the exact dataset and related music-emotion work. WCMED, the Western Classical Music Excerpts Dataset, is a curated corpus of short Western classical music excerpts introduced for a comparative study of whether models developed for soundscape analysis can transfer to music emotion recognition (Fan et al., 2020). It was constructed in parallel with CCMED, the Chinese Classical Music Excerpts Dataset, so that Western and Chinese classical music could be analyzed under the same annotation framework and modeling pipeline. In that original setting, WCMED was not presented as a standalone benchmark for conventional music emotion recognition alone; its primary function was methodological and comparative, as the Western-classical counterpart against which the transferability of soundscape-based models could be measured (Fan et al., 2020).

1. Definition and research role

WCMED was introduced in the study “A Comparative Study of Western and Chinese Classical Music based on Soundscape Models” (Fan et al., 2020). The central research question was whether sound event detection (SED) and soundscape emotion recognition (SER) models could be used to analyze and compare Western and Chinese classical music. Within that design, WCMED served as the Western half of a paired benchmark, while CCMED served as the Chinese half.

This construction gives WCMED a specific epistemic status. It is not merely a corpus of Western classical excerpts with valence–arousal labels. It is a dataset designed to test transferability across domains: from soundscape recordings to Western classical music, and comparatively from soundscape recordings to Chinese classical music. A common misconception is to treat WCMED as just another music-emotion dataset. The original paper instead frames it as a controlled instrument for comparing how far soundscape-derived representations carry into different musical traditions (Fan et al., 2020).

A later study on the data distribution gap in music emotion recognition reused WCMED, abbreviating it as W2, and positioned it as a stylistically distant classical dataset that exposes out-of-distribution weakness in standard MER systems (Ching et al., 6 Oct 2025). That later use reinforces the original paper’s implication that WCMED is especially informative when the question is not only in-domain accuracy, but robustness under distribution shift.

2. Corpus design and audio characteristics

In the original comparative study, WCMED contains 400 excerpts drawn from Western classical music recordings (Fan et al., 2020). The excerpts were selected from the Saarland dataset, described there as “Saarland Music Data (SMD)”, a set of royalty-free audio recordings of 200 pieces and movements from the Western classical music repertoire. The instrumentation was chosen to be diverse in order to cover a wide range of timbres. Instruments represented include piano, violin, viola, cello, double-bass, flutes, trumpets, trombone, and xylophone, among others (Fan et al., 2020).

The excerpt duration was set to 8–20 seconds, following prior recommendations for music mood analysis and with the practical aim of balancing phrase completeness against timbre homogeneity within each excerpt. For WCMED specifically, the average duration is 12.41 seconds with standard deviation 2.74 seconds. All excerpts were converted to WAV, with 44,100 Hz sampling frequency, 32-bit precision, mono channel. No fixed official train/validation/test split was defined at the dataset level in the original paper; evaluation instead relied on repeated random splits during experiments (Fan et al., 2020).

A later cross-dataset MER study describes WCMED as a dataset of 200 royalty-free audio recordings of Western classical repertoire, with 79 of those 200 being solo piano pieces by Bach, Beethoven, Chopin, Mozart, and Rachmaninoff, collected from the Saarland Music Dataset (SMD) (Ching et al., 6 Oct 2025). This later characterization suggests a distinction between the number of underlying source recordings and the number of annotated excerpts used in the original comparative study.

Property	WCMED description
Primary role	Western counterpart to CCMED for soundscape-model transfer
Size in original study	400 excerpts
Source collection	Saarland Music Data (SMD)
Source repertoire description	200 pieces and movements from Western classical music
Excerpt duration	8–20 s
WCMED mean duration	12.41 s
Audio format	WAV, 44,100 Hz, 32-bit, mono

The corpus design is consequential for interpretation. The short clips support timbral consistency and manageable annotation, but they also reduce access to long-range musical structure. This matters particularly because the original authors later argue that Western classical valence depends more heavily on harmony, melody, and orchestration, which are not well captured by soundscape-oriented models (Fan et al., 2020).

3. Annotation methodology and label semantics

WCMED was annotated for perceived emotion, not induced emotion, using the two-dimensional valence–arousal model of Russell (Fan et al., 2020). In the paper’s terminology, arousal is the level of activation and valence is the level of pleasantness. Rather than collecting direct scalar ratings, the authors adopted a ranking-based pairwise comparison protocol, which they state had been used in earlier work and was selected because it simplifies the task and improves inter-annotator reliability (Fan et al., 2020).

WCMED and CCMED were annotated in separate crowdsourcing experiments, and valence and arousal were also collected in separate studies. The platform used was Figure Eight. The ranking procedure was implemented through an iterative Quicksort-based comparison scheme. In the first iteration, one excerpt is chosen as a pivot and all other excerpts are compared against it. For each pairwise comparison, the system collects 3 annotations, and the outcome is decided by majority vote, meaning the selected answer must be chosen by at least two annotators. The procedure then recurses on the resulting subsets until all excerpts are ranked (Fan et al., 2020).

The quality-control pipeline was explicit. Annotators first read a tutorial explaining valence and arousal, then completed a 5-item quiz with hand-selected gold-standard comparisons. They needed 70% accuracy on this quiz to continue. During the main task, additional gold-standard comparisons were inserted for ongoing quality control. Annotators could replay audio excerpts, were required to use headphones, and the interface disabled the volume control bar so that participants could not adjust individual excerpt volumes (Fan et al., 2020).

Across the study as a whole, 989 annotators from 21 countries participated. After ranking, the authors converted rankings into continuous targets for regression by linearly mapping rank positions 1 to 400 into the range 1.0 to -1.0. They explicitly note two assumptions behind this conversion: first, that valence and arousal lie in $[1.0, -1.0]$; second, that successive ranks are equally spaced. This is an important methodological caveat, because ordinal rankings are thereby given interval-scale structure (Fan et al., 2020).

Inter-annotator reliability was assessed using percent agreement and Krippendorff’s alpha. Across the full study, percent agreement was 77.3% for arousal and 76.1% for valence. Krippendorff’s alpha values were between 0.21 and 0.40, which the authors interpret as indicating a fair level of agreement. No separate reliability figure was reported for WCMED alone (Fan et al., 2020).

This annotation scheme later became analytically important in cross-dataset work. The 2025 study explicitly notes that WCMED differs from other MER datasets because it uses pairwise ranking rather than direct scalar ratings, and that the adaptability between pairwise and direct annotation schemes deserves further study (Ching et al., 6 Oct 2025).

4. Transfer-learning experiments from soundscape models

WCMED was used in two principal transfer-learning settings in the original paper: SED + SVR and SER + SVR (Fan et al., 2020).

In the SED-based setup, the model is VGGish, pretrained on AudioSet, which contains 2,084,320 human-labeled 10-second audio clips. VGGish is used as a feature extractor rather than fine-tuned end-to-end. For each WCMED excerpt, it computes log-Mel spectrograms and outputs 128-dimensional embedding vectors. The authors discard the first and last embedding vectors for each audio file because the timbre at the beginning and end of excerpts tends to differ from the middle. The embeddings are then used as inputs to support vector regression (SVR) with an RBF kernel, with C and gamma selected by grid search (Fan et al., 2020).

In the SER-based setup, the source task is emotion prediction on the Emo-soundscapes dataset, which contains 1213 soundscape clips of 6 seconds each with perceived emotion rankings in 2D valence–arousal space. The authors first extract 305-dimensional low-level spectrum features with Essentia, using frame size 2,048 and hop size 1,024. They then apply a windowing-based data augmentation strategy in which 80 frames are selected as one augmented sample, corresponding to about 1.86 seconds at 44.1 kHz. Two LSTM-RNN models are trained on Emo-soundscapes, one for arousal and one for valence. For transfer to WCMED, the final output layer is removed and the penultimate layer activations are used as 128-dimensional embeddings, which are again fed to SVR models (Fan et al., 2020).

The evaluation protocol is the same in both transfer settings: the dataset is shuffled 10 times; on each repetition, 10% is randomly selected for testing and the remaining 90% is used for training. The train/test split is performed before embedding extraction to avoid leakage. Because one excerpt yields multiple embeddings, prediction is done at the embedding level and then aggregated to the clip level with an ensemble average. Evaluation uses $R^2$ and MSE (Fan et al., 2020).

For WCMED, the quantitative results are as follows:

Transfer setup	Target	$R^2$	MSE
SED + SVR	Arousal	0.687	0.098
SED + SVR	Valence	-0.026	0.325
SER + SVR	Arousal	0.421	0.174
SER + SVR	Valence	0.003	0.317

These results establish the asymmetry that defines WCMED’s empirical role in the paper. Arousal is predicted reasonably well, especially with SED + SVR, but valence is not. In the SED setup, the valence score is negative, meaning worse than a constant mean baseline. Relative to CCMED, the WCMED–CCMED difference is not statistically significant for arousal, but is significant for valence, with $p < 0.001$ in the SED setup and $p < 0.01$ in the SER setup (Fan et al., 2020).

The authors attribute the general superiority of SED transfer over SER transfer to the fact that VGGish was pretrained on the much larger and more diverse AudioSet, whereas the SER models were trained on the smaller Emo-soundscapes dataset (Fan et al., 2020). A further notable point is that VGGish was pretrained on a source that includes Western classical music but not Chinese classical music, yet the transferred system still performs better on CCMED than on WCMED. The authors interpret this as evidence that acoustic resemblance to soundscapes matters more than cultural proximity to the pretraining corpus in this setting (Fan et al., 2020).

5. Feature behavior, interpretation, and methodological caveats

The original paper includes a feature-group analysis using SVR with a linear kernel plus recursive feature elimination, examining four feature groups: loudness, rhythm, tonal, and timbre (Fan et al., 2020). For WCMED arousal, the reported $R^2$ values are 0.553 for loudness, 0.395 for rhythm, 0.046 for tonal, and 0.691 for timbre. For WCMED valence, the corresponding values are 0.061, 0.136, 0.152, and 0.085.

These figures show that, within WCMED, timbre is highly informative for arousal, whereas none of the four feature groups is strongly effective for valence (Fan et al., 2020). The authors interpret this asymmetry by arguing that the soundscape-derived embeddings are better aligned with tonal and timbral cues that matter for Chinese classical music and for soundscapes than with the cues governing emotion in Western classical music. They further argue that Western classical valence depends more on high-level musical structure, specifically harmony, melody, and orchestration, which are not well captured by soundscape-oriented SED models and not captured well enough by the transferred SER models either (Fan et al., 2020).

The paper also reports an indirect binary classification experiment distinguishing WCMED vs. CCMED using the same low-level features as in the SER setup and an LSTM-RNN with a sigmoid output neuron, Adam optimizer, learning rate 0.001, and 10% validation split over 100 epochs. The classifier achieves 91.93% binary accuracy. When applied to 3639 soundscape feature matrices from Emo-soundscapes, 3182 (85.97%) are classified as Chinese classical music rather than Western classical music. This result is not itself an emotion-prediction result on WCMED, but the authors use it to support the broader interpretation that soundscape data are acoustically more similar to CCMED than to WCMED (Fan et al., 2020).

Several limitations of WCMED are explicit or strongly implied in the original study. The dataset is small, with 400 excerpts. The labels are derived from rankings mapped linearly to $[-1,1]$, introducing unvalidated interval assumptions. Agreement is only fair by Krippendorff’s alpha. The excerpts are short, which favors timbral homogeneity but may undersample large-scale structural cues. The transfer models were designed for sound events or soundscape affect, not for the symbolic and compositional structures especially relevant to Western classical valence. Finally, there is no fixed public benchmark split, because the paper evaluates using repeated random partitions (Fan et al., 2020).

6. Later reuse in out-of-distribution MER research

WCMED acquired a second major role in the later study “A Study on the Data Distribution Gap in Music Emotion Recognition” (Ching et al., 6 Oct 2025). That paper investigates five datasets with dimensional valence–arousal annotations—EmoMusic, DEAM, PMEmo, WTC, and WCMED—and uses WCMED as a particularly revealing test case for out-of-distribution generalization.

In that study, WCMED is described as a Western classical dataset with 200 royalty-free audio recordings, including 79 solo piano pieces, and with pairwise-ranking annotations ranging from 0 to 400 for both valence and arousal, later normalized independently to $[-1,+1]$ before training and evaluation (Ching et al., 6 Oct 2025). The paper emphasizes four properties that make WCMED distinctive: its Western classical repertoire, its inclusion of solo piano and broader classical excerpts, its pairwise-ranking annotation scheme, and its status as a crucial stress test for cross-dataset generalization (Ching et al., 6 Oct 2025).

The paper’s cross-dataset matrix makes WCMED especially diagnostic. Using Jukebox embeddings + MLP, when WCMED is the test set, the results are:

• train on EmoMusic: Avg. $-0.84$, A $-1.12$, V $R^2$0
• train on DEAM: Avg. $R^2$1, A 0.06, V $R^2$2
• train on PMEmo: Avg. $R^2$3, A $R^2$4, V $R^2$5
• train on WTC: Avg. $R^2$6, A 0.58, V $R^2$7
• train on WCMED itself: Avg. 0.81, A 0.75, V 0.87 (Ching et al., 6 Oct 2025)

This is one of the clearest empirical statements about WCMED in the literature: in-domain prediction on WCMED is good, but out-of-domain transfer to WCMED is very poor (Ching et al., 6 Oct 2025). Even training on WTC, another classical dataset, does not transfer well overall. Arousal transfers somewhat, but valence remains strongly negative.

The later paper then analyzes why. In t-SNE visualizations of Jukebox embeddings, WCMED and WTC occupy extreme positions and remain distinct from each other. The authors explicitly note that the two classical datasets are both extreme yet still different in feature space (Ching et al., 6 Oct 2025). Pairwise divergence calculations reinforce this. For WCMED vs EmoMusic, the reported values are Data WD = 1.71, JS = 0.46, while Annotation WD = 0.14, JS = 0.02. For WCMED vs WTC, Data WD = 1.59, JS = 0.43, while Annotation WD = 0.13, JS = 0.51 (Ching et al., 6 Oct 2025).

These numbers support two important conclusions. First, WCMED is distributionally extreme in data-content space. Second, normalized label distributions can appear deceptively similar even when the underlying musical content differs profoundly. The paper therefore argues that label normalization across datasets may conceal mismatches between music content and emotion semantics (Ching et al., 6 Oct 2025).

The same study also reports that WCMED benefits strongly from a more robust representation and training mixture. In a final OOD evaluation where WCMED is test-only, the best result is obtained with Jukebox + Chroma trained on a combined set of EmoMusic + PMEmo + WTC, yielding Avg. $R^2$8, A $R^2$9, V $R^2$0 on WCMED (Ching et al., 6 Oct 2025). The authors interpret this as evidence that Jukebox embeddings are powerful but dataset- and genre-sensitive, whereas Chroma features are more uniformly distributed across datasets and genres and therefore help stabilize transfer (Ching et al., 6 Oct 2025).

In this later literature, WCMED is thus not only a Western-classical emotion dataset; it becomes a probe dataset for OOD MER. A plausible implication is that WCMED’s enduring value lies less in serving as a high-volume benchmark and more in functioning as a distribution-shift diagnostic for models that appear strong on more homogeneous or more popular-music-oriented corpora.

7. Significance and interpretation

WCMED’s significance derives from the interaction of three properties documented across the two papers. First, it is a paired comparative dataset, explicitly designed to test soundscape-model transfer against a Chinese counterpart under matched annotation and evaluation conditions (Fan et al., 2020). Second, it is a Western classical dataset whose arousal is learnable from timbral and sound-event-oriented representations, but whose valence is not well captured by those representations (Fan et al., 2020). Third, it is a distributionally difficult target domain whose mismatch with other MER datasets remains substantial even when those datasets are also nominally classical (Ching et al., 6 Oct 2025).

These findings make WCMED methodologically valuable for at least two distinct reasons. In transfer-learning research, it exposes the boundary between what soundscape-derived acoustic representations can and cannot recover from Western classical excerpts. In cross-dataset MER research, it reveals the extent to which strong in-domain performance can fail to generalize under stylistic and annotation mismatch.

The broader interpretive lesson is restrained but clear. The original study concludes that soundscape models can analyze Western classical music, because nontrivial arousal prediction is possible, especially with VGGish-based transfer. At the same time, WCMED consistently underperforms CCMED, especially on valence, and this underperformance is one of the paper’s strongest findings (Fan et al., 2020). The later distribution-gap study reaches a complementary conclusion: WCMED should not be assumed interchangeable with WTC or with “classical music” as a generic category, because it remains a distinct domain in both feature space and generalization behavior (Ching et al., 6 Oct 2025).

The dataset has also been publicly shared together with CCMED at http://metacreation.net/ccmed_wcmed_soundscape/ (Fan et al., 2020). That availability, combined with its documented sensitivity to domain shift, has made WCMED a useful resource not only for emotion prediction itself but also for studying how representation learning, annotation design, and genre specificity interact in music-information-retrieval systems.