We describe the architecture of a multiDSP system for large vocabulary speech recognition, called RICO, and we characterize this system both in accuracy and in throughput for 3 subcases (one step isolated words, two steps isolated words,one step continuous speech) on a 1008 words Italian test vocabulary. Then we describe some architectural improvement performed in the following and we characterize the corresponding throughput improvement measured for the most demanding case, i.e. one step continuous speech recognition.