This paper investigates the effectiveness of a speech training application with question answering problems based on speech recognition which is robust to noise in a classroom. In actual classrooms, since a lot of students speak at the same time, input speech waveforms include speech noise from neighboring students. To the best of our knowledge, no speech training study has examined speech recognition with the focus on actual noise corrupted speech input in classrooms. Since existing mobile applications assumed solitary voice input, they failed to evaluate the input as corrupted by speech noise. To maintain students' motivation even in noisy environments, our application ignores the insertion errors around the user's intended sentence. To adapt to junior high school students' speech and the background speech noise, we introduce unsupervised adaption with matched sentences by comparing the speech recognition results and target sentences candidates. We also improve the user interface by reflecting the feedback from teachers and students. The results of a two month trial with over 140 students in a public junior high school show that our speech recognizer improves accuracy and our application achieves a positive user experience.