Scholarlink

<p>Accurate audio segmentation and classification are essential for radio and broadcast analytics, supporting applications like speech recognition, music retrieval, and advertisement monitoring. Radio streams, however, present challenges such as overlapping events, abrupt transitions, and broadcast noise, requiring real-time processing. Traditional rule-based methods struggle with these complexities, while frame-based machine learning models often need inefficient post-processing. This paper reviews current audio analysis techniques, from early methods like short-time energy analysis and spectral features to classical machine learning models such as Hidden Markov Models and Support Vector Machines. More recently, deep learning approaches including convolutional, recurrent, and transformer-based architectures have significantly advanced the field, with models like YAMNet, Spleeter, and Wav2Vec 2.0. We focus on the YOHO algorithm, which reformulates segmentation as boundary regression using computer vision models, offering improved precision and real-time efficiency. The paper also goes through advanced architectures that integrates source separation, YOHO-style regression, and transformer embeddings to achieve precise, timestamped segmentation at broadcast scale. Finally, we discuss future research directions, including noise-robust models, self-supervised learning, multimodal fusion, and real-time solutions.</p>

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

A Comprehensive Review of Audio Segmentation and Classification: from Rule-Based Methods to YOHO and Deep Learning Approaches

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Abid Iqbal,
  • Irfan Ahmed

摘要

Accurate audio segmentation and classification are essential for radio and broadcast analytics, supporting applications like speech recognition, music retrieval, and advertisement monitoring. Radio streams, however, present challenges such as overlapping events, abrupt transitions, and broadcast noise, requiring real-time processing. Traditional rule-based methods struggle with these complexities, while frame-based machine learning models often need inefficient post-processing. This paper reviews current audio analysis techniques, from early methods like short-time energy analysis and spectral features to classical machine learning models such as Hidden Markov Models and Support Vector Machines. More recently, deep learning approaches including convolutional, recurrent, and transformer-based architectures have significantly advanced the field, with models like YAMNet, Spleeter, and Wav2Vec 2.0. We focus on the YOHO algorithm, which reformulates segmentation as boundary regression using computer vision models, offering improved precision and real-time efficiency. The paper also goes through advanced architectures that integrates source separation, YOHO-style regression, and transformer embeddings to achieve precise, timestamped segmentation at broadcast scale. Finally, we discuss future research directions, including noise-robust models, self-supervised learning, multimodal fusion, and real-time solutions.