A combination of data mining method with decision trees building for speech/music discrimination

Qiong Wu, Qin Yan, Jun Wang, Jun HongInterspeech 2008

TL;DR

This paper introduces a novel speech/music discrimination system that combines data mining techniques with hierarchical oblique decision trees. By carefully selecting optimal feature subsets and employing a multi-layered decision structure, the system achieves state-of-the-art accuracy with significantly lower computational complexity, making it ideal for real-time multimedia applications.

What Problem Does This Paper Address?

In the ever-expanding world of multimedia, accurately and efficiently distinguishing between speech and music in audio streams is crucial for a variety of applications, from content-based retrieval to intelligent media management. Existing methods often struggle to balance accuracy with low latency and computational efficiency, which are essential for real-time processing. This paper aims to bridge this gap by developing a robust and computationally inexpensive speech/music classifier.

Methods

The core of this work lies in a two-pronged approach: feature selection and a novel classification architecture.

  • Feature Selection: The authors utilize a Feature Subset Selection (FSS) tool to identify optimal combinations of acoustic features. They recognize that features with different time delays (short-term vs. long-term) have complementary strengths and weaknesses. To address this, they divide features into two groups based on their time-delay characteristics and select optimal subsets for each. Notably, they also incorporate "Harmonic structure stability" (Hss) after observing its effectiveness in music discrimination.
  • Hierarchical Oblique Decision Trees (HODT): Instead of traditional axis-parallel decision trees (like CART), the paper proposes using oblique decision trees (ODTs), specifically employing the OC1 algorithm. ODTs create hyper-plane partitions in the feature space, offering more flexible and potentially more accurate classifications than the axis-aligned partitions of CART. To further enhance performance and handle uncertainty, they implement a double-layer hierarchical DT structure. This structure processes features in two stages, outputting probabilities for music and speech. A "hang-over" strategy, using a music_Cnt counter, is employed to refine uncertain classifications.

Datasets

  • Training Data: The system was trained on a corpus of monophonic speech and music samples, band-limited to 16 kHz and sampled at 100 Hz. This data was sourced from the National 863 Chinese Speech database and ITU-T codec test sets, encompassing various music genres and speech from multiple speakers in English, French, and Chinese. The training set included pure and 20 dB SNR audio without silence, comprising 228,512 music frames and 23,7671 speech frames.
  • Testing Data: An independent test set of 450 music and speech files was used. These files were from the same database but entirely distinct from the training data. They were tested at three SNR levels: pure, 30 dB, and 20 dB, with labels generated every 10 ms. The test files were further categorized into speech-only, music-only, and speech-and-music mixed genres.

Key Insights and Ideas

  • Complementary Feature Sets: The paper highlights the importance of combining short-time and long-time delay features. Neither set alone is sufficient for optimal performance, and their judicious selection is key.
  • Oblique Trees for Flexibility: Oblique decision trees (ODTs) offer a more powerful partitioning capability than axis-parallel trees, allowing for more nuanced discrimination, especially in complex feature spaces.
  • Hierarchical Structure for Robustness: A multi-layered hierarchical decision tree structure effectively handles uncertainty and improves overall classification accuracy by refining decisions across multiple stages.
  • Hss as a Key Discriminator: The "Harmonic structure stability" feature, while having low correlation with some other features, proved to be a valuable addition for improving music discrimination.
  • Low Complexity is Achievable: The proposed system demonstrates that high accuracy can be achieved without sacrificing computational efficiency, a critical factor for real-time applications.

Key Results

The proposed Hierarchical Oblique Decision Tree (HODT) system achieved impressive results:

  • Pure Environment: A frame-level accuracy of 97.6% was reported.
  • High SNR Environment (20dB): An accuracy of 95.2% was achieved.
  • Comparison: The HODT system outperformed the AMR-WB+ codec (used as a benchmark) by a significant margin in both pure and high SNR conditions. It also showed superior performance compared to CART and other oblique decision tree variations.
  • Computational Complexity: The system's complexity is reported to be lower than 1 Weighted Million Operations per Second (WMOPS), making it highly suitable for embedded systems and real-time processing.
  • Frame Rate: The system outputs classifications every 10 ms, enabling fine-grained, segment-level analysis.

Limitations and Future Work

  • Limitations: The paper notes that at lower SNR levels (e.g., 20dB), the system's performance in discriminating music can be affected by noise, leading to some overlap with speech features.
  • Future Work: The authors suggest focusing on improving accuracy in lower SNR environments and extending the algorithm to perform music genre categorization, moving beyond simple speech/music discrimination.

Summary updated

BibTeX
@inproceedings{wu2008combination,
  author = {Qiong Wu and Qin Yan and Jun Wang and Jun Hong},
  title = {{A combination of data mining method with decision trees building for speech/music discrimination}},
  booktitle = {Interspeech},
  year = {2008},
}
Figure