Similarity Measures for Chinese Pop Music Based on Low-level Audio Signal Attributes

Tan Lee 李丹, Chun-Man Mak, Tan Lee, Suman Senapati, Yu Ting Yeung, Wang-Kong LamIsmir 2010

Introduction

As digital music libraries continue to expand, the demand for sophisticated recommendation systems has grown. While metadata-based systems (using artist or genre tags) and collaborative filtering are common, they often fail to surface niche or new content. Content-based recommendation, which analyzes the raw audio signal, offers a solution. However, most research in this field focuses on "cross-genre" classification (e.g., Jazz vs. Rock). This study addresses the more difficult challenge of intra-genre similarity within Chinese pop music, investigating whether low-level audio descriptors can effectively model human perception of musical similarity.

Methodology and System Architecture

The researchers proposed a framework to decompose songs into a multi-dimensional attribute vector. The process begins with signal segmentation and proceeds through several specialized analysis modules.

System architecture for extracting the song attribute vector from a raw audio signal.

1. Vocal/Non-vocal Segmentation

Using Mel-frequency cepstral coefficients (MFCCs) and Gaussian Mixture Models (GMMs), the system first distinguishes between instrumental and singing segments. This stage achieved a high accuracy of 90%, providing a clean foundation for subsequent analysis.

2. Feature Extraction

  • Instrument Identity: Non-vocal segments are analyzed to identify the presence of eight common instruments, such as piano, strings, and percussion.
  • Singing Style & Gender: Vocal segments are compared against 13 reference classes based on prominent Chinese pop stars to determine style and gender.
  • Tempo: Extracted via Fourier analysis of beat onset patterns to generate a "tempogram," capturing the rhythmic pulse of the track.
  • Degree of Noisiness (DoN): This novel metric combines signal intensity and spectral flatness to capture the "thickness" of the arrangement.

Modeling Emotion through Noisiness

A key innovation of this research is using the "Degree of Noisiness" (DoN) as a proxy for high-level emotional mood. The authors mapped these low-level acoustic features to Thayer’s emotion model, which plots emotions along axes of Energy and Stress.

Thayer’s emotion model used to map 'Degree of Noisiness' to musical mood.

In this model, "thick" textures and high intensity (high DoN) correlate with "excited" or "angry" moods, while "thin" textures (low DoN) correlate with "calm" or "relaxed" vibes. By calculating DoN separately for the first and second halves of a song, the system accounts for the dynamic shifts common in pop music (e.g., a quiet verse leading to a loud chorus).

Experimental Results

The system was evaluated using a dataset of 43 query songs and subjective ratings from 215 human listeners. Each listener rated the similarity between a query song and ten candidates on a scale from "very different" to "very similar."

Attribute Average Correlation to Human Scores
Degree of Noisiness 0.49
Tempo 0.24
Singing Style 0.20
Instrument 0.07
Gender 0.06

The results revealed a clear hierarchy: listeners prioritize the "vibe" (mood and rhythm) over technical specifications like the singer's gender or specific instrumentation. By weighting DoN and Tempo most heavily, the system achieved an overall average correlation of 0.544 with human judgment.

Conclusion

This study demonstrates that low-level audio features, particularly those capturing energy and texture, can provide a "fairly good" approximation of how humans perceive similarity within a single genre. While the model currently lacks high-level semantic features like melody or lyrics, the success of the "Degree of Noisiness" metric suggests that capturing the emotional atmosphere of a song is a critical component for future content-based recommendation engines.

Summary updated

BibTeX
@inproceedings{2010similarity,
  author = {Tan Lee 李丹 and Chun-Man Mak and Tan Lee and Suman Senapati and Yu Ting Yeung and Wang-Kong Lam},
  title = {{Similarity Measures for Chinese Pop Music Based on Low-level Audio Signal Attributes}},
  booktitle = {Ismir},
  year = {2010},
}
Figure