Scholarlink

Recent advancements in deep learning have revolutionized audio source separation, addressing challenges like the cocktail party problem. This study explores the efficacy of Convolutional Neural Networks (CNNs), specifically the Conv-TasNet architecture, for separating vocals, drums, bass, and other instruments from mixed audio using spectrogram-based methods, and deep learning models. Leveraging the MUSDB18 dataset, these models were trained and evaluated, and the best model achieved significant performance metrics such as an accuracy of 0.93, Signal-to-Distortion Ratio (SDR) of 13.2 dB, Signal-to-Interference Ratio (SIR) of 14.6 dB, and Signal-to-Artifact Ratio (SAR) of 10.5 dB. Additionally, explainability techniques like SHAP and Integrated Gradient were employed to enhance model interpretability. These findings contribute to advancing more efficient and interpretable audio processing systems.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable Machine Learning Techniques for Audio Source Separation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Joan Kaitesi,
  • Ssenkungu Reinol Martin,
  • Sudi Murindanyi,
  • Paddy Junior Asiimwe,
  • Benjamin Kanagwa,
  • Ggaliwango Marvin

摘要

Recent advancements in deep learning have revolutionized audio source separation, addressing challenges like the cocktail party problem. This study explores the efficacy of Convolutional Neural Networks (CNNs), specifically the Conv-TasNet architecture, for separating vocals, drums, bass, and other instruments from mixed audio using spectrogram-based methods, and deep learning models. Leveraging the MUSDB18 dataset, these models were trained and evaluated, and the best model achieved significant performance metrics such as an accuracy of 0.93, Signal-to-Distortion Ratio (SDR) of 13.2 dB, Signal-to-Interference Ratio (SIR) of 14.6 dB, and Signal-to-Artifact Ratio (SAR) of 10.5 dB. Additionally, explainability techniques like SHAP and Integrated Gradient were employed to enhance model interpretability. These findings contribute to advancing more efficient and interpretable audio processing systems.