顧客關係管理為協助企業能更有效的管控顧客，並執行出正確的決策以拉近與顧客之間的關係，使該企業的顧客保留率上升，達成企業利潤提升的結果。然而由Satisfcation- Profit Chain得出影響到顧客保留率的重要因子為顧客滿意度，顧客滿意度是顧客對於產品的期望與實際使用過價值之間的落差，為最貼近顧客真實想法及感受的指標，為此顧客滿意度成為企業炙手可熱的績效指標。縱然，搜集顧客滿意度方法有許多，但本研究認為透過顧客的聲音能直接判別顧客最真實的滿意程度，恰巧聲音被譽為人類與生俱有的溝通方式，為此得知聲音是能傳遞資訊的重要工具，本研究將建立一個聲音辨識滿意度的模型，透過梅爾倒頻譜係數萃取顧客聲音中滿意與不滿意的特徵值，再將獲得的聲音特徵以卷積神經網路訓練，以分類顧客滿意與否的模型。 與此同時，在建立完聲音辨識滿意度的模型後，本研究將優化聲音辨識滿意度模型的參數使模型更精進以發揮出極佳的效能，因此調整參數至適配程度最高也是本研究致力於之，包括模型使用的兩種方法：（1）調整梅爾倒頻譜係數中萃取聲音的重要參數，使梅爾倒頻譜係數萃取出更好的特徵值，（2）利用卷積神經網路使萃取出的聲音特徵，進行有意義的特徵擷取和分類，以測出本研究的最佳參數組合，讓聲音辨識滿意度模型的穩定性及辨識性紛紛提升。;Customer Relationship Management（CRM）helps companies manage their customers more effectively and implement correct decisions. Then Companies try to have more relationships with their customers, increasing the company′s customer retention rate, and improving the company′s profits. However, the important factor in the Satisfaction- Profit Chain influences the customer retention rate obtained by customer satisfaction. Customer satisfaction is the gap between the customer’s expectations of the product and the actual used value, which is the closest performance indicator. Customer satisfaction has become an important performance indicator for companies. Although there are many ways to collect customer satisfaction. This study believes the customer’s voice can directly determine the true satisfaction level of customers. Vocal is hailed as the innate communication method of human beings. For this reason, it is known that vocals can convey important information. This research will establish a vocal satisfaction identification model that extracts the satisfied and dissatisfied features of the customer’s voice through the Mel-Frequency Cepstral Coefficients（MFCC）. Next, classify the customer’s satisfied and dissatisfied features by Convolutional Neural Network（CNN）. At the same time, the establishment of the vocal satisfaction identification model will optimize the parameters. To make the model more refined and achieve more excellent performance. For this reason, the optimization parameters by the model: （1） Adjust the important parameters of the extracted sound in the MFCC, so that the MFCC can extract better eigenvalues, （2） Use the CNN method to make the extracted sound features meaningful and classified. By experiment with the best parameter of this research, getting a vocal satisfaction identification model is more stable and recognizable.
