智慧型家用機器人使用之語者辨識系統

Abstract

本研究探討語者辨識系統於智慧型家用機器人上以及家用環境內，所會面臨的各式使用情境，並針對情境開發設計我們的語者辨識系統及註冊流程，以期使用者能更方便直覺地使用本系統。此外，因為家用環境中通常富含各式雜訊，且使用者在使用機器人時必定與機器人有一段距離，所以將文本獨立語者辨識前端整合麥克風陣列、波束形成和空間濾波器的技術，以提高在雜訊環境下的強健性。另一方面也考慮到在家用機器人中，與其他系統整合的可能性，在系統輸出端則提供辨識結果有效與否的判定及信心指數，以利與其他使用者辨識系統整合。最後為了更貼近家庭使用者的使用習慣與喜好，並降低所需註冊及測試語料的秒數，而發展出可用任何語言輸入註冊文本資訊及註冊語料的文本相關語者辨識系統。

In this paper we present a speaker recognition system, which is specifically designed for intelligent home robot. The enrollment procedure of the system is designed to fit home scenarios and make it easier to be used. Besides, the performance of speaker recognition system degrades significantly in home environment because of reverberation, noise and the distant between speakers and microphone array. The spatial information from the microphone array, which couples with beam forming and spatial voice activity detection, makes the system more robust in the noisy environment. On the other hand, our system provides the confidence scores so as to be fused with other recognition results and achieve the integration of user recognition systems. Finally, in order to increase the convenience of using systems and reduce requirements of enrollment and test data, we develop a text-dependent speaker recognition system, which can be used in any language.