Skip to content
Open access

Dimension-Wise Probing of Music Foundation Models for Chinese and Western Music: Timbre, Technique, Key, and Mode

Aug 2026 · Electronics · Vol 15, pp. 3421 · 0 citations · 10 references

TL;DR

These results map which musical dimensions are readily decodable from the evaluated frozen representations while avoiding a cultural-capability interpretation of the unmatched cross-corpus comparison.

Abstract

Music foundation models are increasingly used as feature extractors in music information retrieval; yet, their treatment of Chinese tonal organization has received little controlled evaluation. We probe frozen MERT, MuQ, CultureMERT, and an 88-dimensional librosa baseline on instrument timbre, guzheng playing technique, Western key, and Chinese pentatonic mode. Probes are evaluated with task-specific splits that guard against leakage and with chance-normalized accuracy. Timbre is near ceiling (0.994–0.998), whereas source-aware guzheng-technique accuracy ranges from 0.382 to 0.787. In the unmatched external comparison, GiantSteps key scores 0.441–0.604 and CNPM mode pattern scores 0.294–0.355; this difference is descriptive because the corpora differ in genre, production, provenance, and label distribution. Same-audio CNPM controls identify an important boundary: the TongGong system reaches 0.649–0.734, and absolute tonic reaches 0.433–0.480, whereas their matched five-way mode-pattern labels reach 0.256–0.350 and 0.243–0.366, respectively. Chroma ablations show that pitch-class energy is central to the external key task but does not by itself account for CNPM mode-pattern performance. These results map which musical dimensions are readily decodable from the evaluated frozen representations while avoiding a cultural-capability interpretation of the unmatched cross-corpus comparison.

Read PDF

Similar papers

#artificial intelligence Preprint Oct 2026

Gugak-VocalSet: A Multi-Genre Studio-Quality Dataset of Korean Traditional Singing with Sigimsae Annotations

We present Gugak-VocalSet, the first publicly available audio dataset for Korean traditional singing with dense annotations of sigimsae, a structured pitch contour system of vocal expressions. Gugak-VocalSet contains 24.5 hours of unaccompanied solo vocal recordings across 305 tracks of representative repertoire curate...

Danbinaerin Han, Wonil Kim, Se-Ah Hong et al. · 0 citations
Open access Sep 2026

Traditional Chinese Music Goes Global: Encoding and Digital Audience Reception in Black Myth: Wukong

Video games can circulate culturally specific music widely, but positive audience response does not by itself show how that music is interpreted. This study examines the encoding and English-language online reception of traditional Chinese musical material in Black Myth: Wukong. The encoding analysis draws on identifia...

Tian-Guang Xu · 0 citations
#machine learning Preprint Sep 2026

A Repeated-Measurement Study for Cultural Analytics of English Song Lyrics Using Five Large Language Models

The findings show that LLM-based measurement is not uniformly reliable across constructs, and Repeated-measurement stability and cross-model convergence should be reported before LLM annotations are treated as scalable measurements in cultural analytics.

E. Smith, S. Ho, D. Laux · 0 citations
#machine learning Preprint Sep 2026

TrueMuse: A Benchmark for Data Attribution in Text-to-Music Models

Text-to-music generation models are trained on massive music collections, creating a growing need for data attribution methods that can quantify the contribution of individual training samples. However, existing attribution methods are difficult to rigorously evaluate due to the lack of reliable ground truth, making it...

Jia-Wei Yu, Jian Liu · 0 citations
#natural language process... Preprint Sep 2026

Don't CLAP: Are Music-Text Models Bag-of-Words?

The results provide compelling evidence that the CLAP score and related metrics do not capture fine-grained musical meaning or attribute bindings; their representation is closer to a bag-of-words that leaves them insensitive to meaning-changing perturbations of the caption.

Yuan-Chiao Cheng, Alexander Lerch · 0 citations

YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality

Symbolic models make melody, harmony, rhythm, and form explicit but typically stop before a finished recording; audio models produce complete songs while leaving composition implicit. We introduce YuE2, which unifies symbolic and audio music generation at frontier quality through symbolic planning. A single AR-NAR Mixt...

Rui-Bin Yuan, Jia-Hao Pan, Jun-Yan Jiang et al. · 1 citation

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.