当前位置: X-MOL 学术Sādhanā › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
A strong hybrid AdaBoost classification algorithm for speaker recognition
Sādhanā ( IF 1.4 ) Pub Date : 2021-07-09 , DOI: 10.1007/s12046-021-01649-6
V Karthikeyan 1 , S Suja Priyadharsini 2
Affiliation  

Recognizing the person from a sample of their voice print or speech samples is known as Speaker recognition. It is an emerging technology, in recent years machine learning (ML) based classification schemes have been observed as one kind of alternate solution for speaker recognition. In this paper adaptive boosting (AdaBoost) combined with a powerful ML classifier (Random Forest) is proposed to handle multi-class imbalanced speaker data classification. Based on weights AdaBoost integrates several sub-classifiers and constructs a robust classifier. A new strong and more accurate technique is proposed by employing Random Forests as the initial stage classifier and AdaBoost as the subsequent stage classifiers, to decide the class that the speaker sample belongs to. Three dissimilar datasets are utilized to estimate the robustness of the proposed hybrid AdaBoost technique. The classification results of the hybrid RF-AdaBoost are evaluated against other state-of-the-art algorithms (kNN, SVM, RF, kNN- AdaBoost, and SVM- AdaBoost), experimental results convey the proposed algorithm improves accuracy as well as stability for the imbalanced speaker data. The f1_score for RF-AdaBoost is 92%, as well as it produces minimum root mean squared value. The stability of the hybrid algorithm is evaluated using Matthews correlation coefficient (MCC), g-means metric value, and variance, it shows RF- AdaBoost outperforms the other state-of-the-art algorithms in all the aspects of speaker recognition.



中文翻译:

一种用于说话人识别的强混合 AdaBoost 分类算法

从他们的声纹或语音样本样本中识别这个人被称为说话人识别。这是一项新兴技术,近年来,基于机器学习 (ML) 的分类方案已被视为说话人识别的一种替代解决方案。在本文中,提出了自适应提升 (AdaBoost) 结合强大的 ML 分类器(随机森林)来处理多类不平衡说话人数据分类。基于权重 AdaBoost 集成了几个子分类器并构建了一个鲁棒的分类器。通过使用随机森林作为初始阶段分类器和 AdaBoost 作为后续阶段分类器,提出了一种新的更强大和更准确的技术,以确定说话者样本所属的类。三个不同的数据集被用来估计所提出的混合 AdaBoost 技术的鲁棒性。混合 RF-AdaBoost 的分类结果针对其他最先进的算法(kNN、SVM、RF、kNN-AdaBoost 和 SVM-AdaBoost)进行评估,实验结果表明所提出的算法提高了准确性和稳定性对于不平衡的说话者数据。RF-AdaBoost 的 f1_score 为 92%,并且它产生最小均方根值。使用 Matthews 相关系数 (MCC)、g-means 度量值和方差评估混合算法的稳定性,它表明 RF-AdaBoost 在说话人识别的所有方面都优于其他最先进的算法。混合 RF-AdaBoost 的分类结果针对其他最先进的算法(kNN、SVM、RF、kNN-AdaBoost 和 SVM-AdaBoost)进行评估,实验结果表明所提出的算法提高了准确性和稳定性对于不平衡的说话者数据。RF-AdaBoost 的 f1_score 为 92%,并且它产生最小均方根值。使用 Matthews 相关系数 (MCC)、g-means 度量值和方差评估混合算法的稳定性,它表明 RF-AdaBoost 在说话人识别的所有方面都优于其他最先进的算法。混合 RF-AdaBoost 的分类结果针对其他最先进的算法(kNN、SVM、RF、kNN-AdaBoost 和 SVM-AdaBoost)进行评估,实验结果表明所提出的算法提高了准确性和稳定性对于不平衡的说话者数据。RF-AdaBoost 的 f1_score 为 92%,并且它产生最小均方根值。使用 Matthews 相关系数 (MCC)、g-means 度量值和方差评估混合算法的稳定性,它表明 RF-AdaBoost 在说话人识别的所有方面都优于其他最先进的算法。以及它产生最小均方根值。使用 Matthews 相关系数 (MCC)、g-means 度量值和方差评估混合算法的稳定性,它表明 RF-AdaBoost 在说话人识别的所有方面都优于其他最先进的算法。以及它产生最小均方根值。使用 Matthews 相关系数 (MCC)、g-means 度量值和方差评估混合算法的稳定性,它表明 RF-AdaBoost 在说话人识别的所有方面都优于其他最先进的算法。

更新日期:2021-07-09
down
wechat
bug