Scholarlink

<p>It is essential to develop a system for identifying the type of bird from the sounds they are chirping. This paper presents sound-based bird classification, and the system is designed based on the spectrogram features and deep learning convolutional neural networks (CNN). Two-dimensional spectrogram features with filters calibrated in BARK and MEL scales and the Gammatonegram features are derived from the sound signals of the birds. These extracted two-dimensional image-like features are applied to the CNN layered architecture, and the trained models are created for each bird. Testing is done by using the spectrogram and Gammatonegram features derived from the test sound on the models. Based on matching the test features and models, the test frame is identified to be associated with the bird. Decision-level fusion of indices corresponding to the correct bird species classification for a CNN-based classifier has provided the maximum 100% accuracy for identifying 20 birds. This work is extended to classify 20 birds spanning different regions using perceptual features and GMM &amp; MHMM modelling techniques, and the decision-level fusion of correct indices of the structures &amp; GMM and MHMM models has provided 99.9% accuracy. This automated bird classification system would help ornithologists determine the ecosystem's health.</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

Sound-based bird classification using multiple features and machine learning paradigms

  • 期刊论文

  • 发表日期：

    2025年1月17日

查看全文
  • Revathi A,
  • Sasikaladevi N

摘要

It is essential to develop a system for identifying the type of bird from the sounds they are chirping. This paper presents sound-based bird classification, and the system is designed based on the spectrogram features and deep learning convolutional neural networks (CNN). Two-dimensional spectrogram features with filters calibrated in BARK and MEL scales and the Gammatonegram features are derived from the sound signals of the birds. These extracted two-dimensional image-like features are applied to the CNN layered architecture, and the trained models are created for each bird. Testing is done by using the spectrogram and Gammatonegram features derived from the test sound on the models. Based on matching the test features and models, the test frame is identified to be associated with the bird. Decision-level fusion of indices corresponding to the correct bird species classification for a CNN-based classifier has provided the maximum 100% accuracy for identifying 20 birds. This work is extended to classify 20 birds spanning different regions using perceptual features and GMM & MHMM modelling techniques, and the decision-level fusion of correct indices of the structures & GMM and MHMM models has provided 99.9% accuracy. This automated bird classification system would help ornithologists determine the ecosystem's health.