In this paper, the authors present the performance of a large vocabulary Japanese spoken sentence recognition system tested under several noise levels (S/N (Signal to Noise Ratio): lOdB, 20dB, 25dB, 30dB, ), and discuss the ability of the system to recover input sentences. The result of the performance evaluation was that the system recovered over 90% of the test sentences when the noise level was low ( °°, 30dB), but recovered only 50% of the test sentences when the noise level reached to an S/N of 20dB.