Training in virtual reality (VR) using a head-mounted display (HMD) has been applied to various fields, and recent attention has been focused on real-time feedback to the trainee. However, trainees' acceptance of the presented feedback is dependent on the state of the ongoing task as well as their internal state. If the trainees are not ready to accept the feedback from the system, it may potentially disturb the ongoing training. This study is premised on it being desirable for a VR training system to recognize the trainee's state and present feedback at an appropriate time. As a first step, this study focused on using a case scenario about public speaking training to confirm the appropriate timing of feedback in terms of trainees' subjective impressions. Specifically, a presentation training system that uses VR-HMD to collect observable multimodal information (including body posture and movement, gaze, heartbeat, perspiration, voice, and slide operation) was developed, which provided real-time feedback using icons. We collected the trainee's subjective scores of disturbance for each feedback in various situations through a Wizard of Oz experiment. Data analysis revealed two things; (1) Real-time feedback negatively affects the trainee's presentation in as many as 30 % of cases,(2) More than half of such negating-impact feedback can be estimated from observable information. This finding is expected to contribute to the development of future training systems that provide real-time feedback avoiding inappropriate timing for the trainee.