根據客戶消費行為研究，客戶滿意度明顯影響客戶回購意願，產品銷售業績和企業成長。通常，滿意度測量需要客戶花費額外的時間來填寫購買後的問卷調查。然而，在電話行銷產業，問卷調查仍存在一些執行上的限制，因此，開發一種更方便收集數據的方法及無需客戶購買後填寫問卷調查即可直接評估客戶滿意度的工具，是非常值得研究。 這項研究試圖驗證，透過語音數據分析是否可以用來區分客戶滿意度。本研究完成了以下任務：設計一個實驗流程來收集客戶表達滿意度的聲音和可驗證相應此聲音的真實數據;根據收集到的數據，將參與者分為滿意、中立和不滿意;從客戶表達滿意度的聲音數據中提取MFCCs（梅爾頻率倒譜係數）作為特徵;由於所收集到的聲音數據集有限，使用Auto-Encoder (自動編碼器) 進一步減少聲音的向量維度;所提取的MFCCs和音韻特徵被輸入到LSTM-RNNs（長期短期記憶-遞歸循環神經網絡）和SVM (支持向量機)，以建立區分客戶滿意度的模型;使用nested cross-validation (雙迴圈交叉驗證) 訓練和評估模型;SVM 和LSTM 的平均準確率分別可以達到 73.97% 和 71.95%。;Customer repurchase behavior is manifestly influenced by satisfaction and the degree of customer satisfaction impacts on business growth and enterprise performance. Typically, satisfaction measurement requires customers to spend additional time to fill in a post-purchase survey. A more convenient way to collect data is to ask customers to express their degree of satisfaction while they are actually using the product or service they have purchased. This study strived to verify if voices can be used to distinguish customer satisfaction. An experiment was set up to collect voices and survey of satisfaction right after participants consume beverage offered. Participants were clustered into satisfaction, neutrality, and dissatisfaction according to the collected questionnaires. The MFCCs (Mel Frequency Cepstral Coefficients) were extracted from voices as features. Due to the dataset was limited, Auto-Encoder was used to reduce the voice features. The extracted and prosodic features were fed into an LSTM-RNNs (Long Short-Term Memory-Recurrent Neural Networks) and SVM models to distinguish customer satisfaction. The nested cross-validation method was used to train and evaluate models. The average accuracy of SVM and LSTM could achieve 73.97% and 71.95% of accuracy, respectively.
