The recent proliferation of API hosting frameworks has dramatically eased the development of interesting web mashups and provided monetization opportunities for enterprises offering high value APIs. Most of these mashups are based on request/response REST model that is widely used in the web world. However, REST is not the primary vehicle for communication oriented services such as Internet telephony, chat, presence or live video communications. Session Initiation Protocol (SIP) is commonly used for communication services and has well defined Java APIs for use by applications and application developers. In this paper, we present 'Talking Cloud' - a API hosting platform on the cloud, for enabling the composition of asynchronous, communication-oriented services to create high value mashups such as object detection in real-time video. Our platform is based on a unique combination of HTTP and SIP, and provides a number of useful features such as service customization, media signaling and routing, social network interactions, elasticity and scaling, besides standard API management functionalities such as usage metering. Our platform provides a flexible framework that allows applications to place hooks within the service invocation workflow, and invoke application logic that is executed remotely, enabling flexible and context sensitive handling of calls/media. We describe our platform architecture and show how communication mashups can interact with it using a combination of REST and SIP interfaces that are exposed by the platform. We describe a prototype implementation and present two usecases - nearest available helpdesk agent and real time object detection within live video streams.