Scholarlink

In this work and the literature, voice signals can be classified as periodic (type 1) or either some periodicity (type 2) and chaos (type 3). This work aims to classify signs into types 1, 2 or 3 to be subsequently applied in a classification system for pathological/control signs. The original dataset is composed of 466 type 1 individuals, 900 type 2 individuals, and 84 type 3 individuals classified by an otolaryngologist. 15% of the data was used for testing and the remaining 85% was used for training and validation. A data augmentation technique was applied to balance the data in training set. Therefore, for the test set, 3380 sounds were used, 1020 type 1, 1280 type 2 and 1080 type 3. Of these, 80% were used for training and 20% for validation. The Mel spectrograms of the signals were used in the input of a VGGish to retrain the model in classifying the 3 types of signals. Regarding test accuracy, this network obtained 71.2%.

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

Nonperiodic Pathologic Voice Signals Classification Using Mel-Spectrogram and VGGish

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Joana Filipa Teixeira Fernandes,
  • João Viana Pinto,
  • Carla Pinto Moura,
  • Helena Vilarinho,
  • Felipe Teixeira,
  • Diamantino Freitas,
  • João Paulo Teixeira

摘要

In this work and the literature, voice signals can be classified as periodic (type 1) or either some periodicity (type 2) and chaos (type 3). This work aims to classify signs into types 1, 2 or 3 to be subsequently applied in a classification system for pathological/control signs. The original dataset is composed of 466 type 1 individuals, 900 type 2 individuals, and 84 type 3 individuals classified by an otolaryngologist. 15% of the data was used for testing and the remaining 85% was used for training and validation. A data augmentation technique was applied to balance the data in training set. Therefore, for the test set, 3380 sounds were used, 1020 type 1, 1280 type 2 and 1080 type 3. Of these, 80% were used for training and 20% for validation. The Mel spectrograms of the signals were used in the input of a VGGish to retrain the model in classifying the 3 types of signals. Regarding test accuracy, this network obtained 71.2%.