Voice is the preferred mode of human communication. Our mostly homogeneous network for voice communications, the Public Switched Telephone Network, is quickly being replaced with a heterogeneous collection of wired and wireless links, including the wired Internet for voice over IP, digital cellular telephone systems, and voice over wireless access points. These links/networks are not jointly designed or standardized, and use different speech coders, so when they are interconnected, there is a significant loss of voice quality. Additionally, we are not well-positioned to take advantage of evolving networks, such as mesh and ad hoc networks, for voice communications. This research develops new speech coders that produce high quality, highly intelligible speech over a wide range of bit rates, with a full range of functionalities, and which are suitable for voice over IP, digital cellular, voice over wireless access points, and woice over wireless ad hoc/mesh networks. The resulting speech coders are especially important for the wireless networks deployed in emergency response situations.
By combining phonetic mode detection and rate distortion theory motivated structures with backward adaptive tree codes, this research derives and develops new speech coders that are SNR and bandwidth scalable, have multiple descriptions functionality, tandem well with other codecs, and have excellent packet loss concealment. The resulting speech coders are important for ad hoc and mesh networks deployed in emergency response situations, where multicast transmissions are prevalent, SNR scalability is important to allow selective pruning of traffic, and multiple descriptions functionality is critical to take advantage of path diversity and provide enhanced reliability. This research establishes new fundamental limits on speech coder performance by combining multimodal speech models and conditional rate distortion theory.
