基於卷積神經網路的情緒語音分析;Emotional Speech Analysis Based on Convolutional Neural Networks

NCU Institutional Repository > 資訊電機學院 > 資訊工程學系碩士在職專班 > 博碩士論文 > Item 987654321/86506

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/86506

題名:	基於卷積神經網路的情緒語音分析;Emotional Speech Analysis Based on Convolutional Neural Networks
作者:	高儀津;KAO, YI CHIN
貢獻者:	資訊工程學系在職專班
關鍵詞:	卷積神經網路;語音偵測;情緒分類;Convolutional Neural Network(CNN);Speech detection;Emotion classification
日期:	2021-10-22
上傳時間:	2021-12-07 12:54:48 (UTC+8)
出版者:	國立中央大學
摘要:	在近期研究中，語音情感識別（speech emotion recognition）已成為人類行為分析研究領域中一個有趣且具有挑戰性的項目。該研究領域的目標是根據人類的語音音調對人們的情緒狀態進行分類。目前，該研究領域的注重於識別語音情緒自動分類器的有效性，以提高實際應用中的分類效率，例如：在電信服務中使用的分類效率，辨識正面情緒（如：快樂、驚訝）和負面情緒（如：悲傷、憤怒、厭惡和恐懼），可以為電信服務的平台用戶和客戶提供大量有效的數據。在本論文中通過使用深度學習技術研究了涉及識別人類語音數據中的正面和負面情緒的複雜任務。本論文中使用了五個開放的情感語音數據集和四個自製語音數據集，分別訓練了多階層的模型來處理正負向的情緒辨識，該模型為正、負面情感語音數據的使用提供了良好的效果。此外，也運用自製數據集進行預訓練模型(七類情緒辨識模型)作為網路參數初始化和從頭開始訓練(Train from scratch)隨機初始化網路參數的方式去比較兩者在做三種族群的語音偵測分類。根據實驗結果，本論文中不論是訓練三種族群的語音偵測分類或者是七類語音的情緒偵測分類，此二種任務便是效能最好的模型都是預訓練模型，可以顯著得看出優於從頭開始訓練(Train from scratch)的模型。;In recent studies, speech emotion recognition has become an interesting and challenging area of research in human behavior analysis. The goal of this research area is to classify people′s emotional states based on their speech tones. Currently, the research area focuses on identifying the effectiveness of automatic classifiers of speech emotions to improve the classification efficiency in practical applications, e.g., for use in telecommunication services, identifying positive emotions (e.g., happiness, surprise) and negative emotions (e.g., sadness, anger, disgust, and fear), which can provide a large amount of valid data for platform users and customers of telecommunication services. In this paper, the complex task of identifying positive and negative emotions in human voice data is investigated by using deep learning techniques. Five open sentiment speech datasets and four self-generated speech datasets are used to train multi-level models for positive and negative sentiment recognition, which provide good results for both positive and negative sentiment speech data. In addition, a pre-trained model (seven types of emotion recognition models) was used to initialize the network parameters, and a random initialization of the network parameters by Train from scratch to compare the two is doing the classification of three groups of speech detection. According to the experimental results, in this paper The best model for both tasks is the pre-training model, which is significantly better than the Train from a scratch model.
顯示於類別:	[資訊工程學系碩士在職專班 ] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	151	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....