In thie paper, we proposes a spontaneous speech recognition method which effectively suppresses false alarms and deletions that are serious problems for speech translation. Our method ("Generation Driven Speech Recognition (GD-SR)") generates sentence hypotheses from speech recognition results and executes speech recognition again using these hypotheses as new language constraints. In the experiment, the proposed method improved 6% in word accuracy and 21% in Japanese-English translation quality in comparison with a conventional N-gram based speech recognizer.