Distributed Speech Recognition

Distributed Speech Recognition (D.S.R.) is based on the idea of decoupling the front-end processing from the rest recognition mechanism using a client-server model over a data network. This way one can use the front-end processing on a lightweight device such as a PDA and 've access to speech information retrieval services using a back-end recognition server. The data network is not restricted to be wire-line. For example, GPRS will "data enable" mobile networks making these services available to GPRS-enabled devices like tomorrow's mobile phones. Hold on! There is more! To make this technology even most appealing to end user it should be cost effective too! To accomplish this the stream of acoustic information transmitted from the client to the server needs to be compacted as much as possible. To achieve this we designed a speech coding scheme specifically for recognition purposes and lowered the data rate to just 2Kbps!

The value of D.S.R. is that it provides substantial recognition performance advantages compared to a conventional mobile voice channel where both the codec compression and channel errors degrade performance. It also enables new mobile multimodal interfaces by allowing the features to be sent simultaneously to other information on a single mobile data channel such as GPRS.