同儕網路(Peer-to-Peer Network)是近年來最熱門的研究領域之一。在此篇論文中，我們利用同儕網路設計了一個互動式的網路電台，其名稱叫做PeerIIR。有別於傳統的網路電台多半是單向的傳輸系統，我們結合了網路電話的技術實現了可以讓許多人互動的網路電台。電台主持人可以邀請其他的共同主持人來一同主持節目。而收聽節目的聽眾，也可以call-in給節目主持人。 然而，在主持人、共同主持人或call-in的人同時講話的時候，會同時產生多條聲音串流來佔用網路頻寬。為了解決這個問題，在節目的主持人只有一個人時，我們會建立樹狀的拓樸來傳送我們的封包。然而，當有多個主持人或是有人Call-in給節目主持人時，我們會透過distributed mixer negotiation的演算法, 在我們所設置的PeerIIR伺服器之間建立樹狀的架構,以分散式地合成聲音串流，達到節省網路頻寬的目的。最後也透過了一連串的實驗模擬來證明我們所設計的系統確實能夠在反應延遲、頻寬負荷上達到好的效能。 / Peer-to-Peer (P2P) applications are popular recently and have become one of the hottest research topics. The participants share their resources (such as processing power, disk storage, and network bandwidth) in the P2P architecture. In this thesis, we design and implement an Interactive Internet Radio system using the P2P approach, called PeerIIR. Different from the traditional radio system which can just broadcast the voice streams to its audience, the PeerIIR utilizes the Voice-over-Internet-Protocol (VoIP) features to make the live interaction between the program host and the audience possible. The program host can invite co-hosts to talk in the program. The audience can also use call-in service to talk with the hosts. However, when the host, co-hosts, and calliners are speaking at the same time, they will produce multiple streams which need to deliver to all the audience on the system. This could consume the network bandwidth inefficiently, or even exhaust the link capacity of the audience. Thus, how to process multiple streams produced at the same time and to deliver to all the audience efficiently is the key issue. When there are only one program host producing the audio stream, a distribution tree is built to distribute it. If there are co-hosts or calliners speaking, a distributed mixer negotiation algorithm is performed to build a voice mixing tree among PeerIIR servers. Therefore the audio streams are mixed distributedly and step by step along the mixing tree to save transmission bandwidth. The results from series of simulation show that the performance for response time and link/node stress is enhanced compared with some related works.