© Springer Nature Switzerland AG 2019. During the feature extraction process for speech recognition, a window function is first applied to the input waveform to extract temporally-limited spectrum. By shifting the window function with a short time period, we can analyze the temporal change of speech spectrum. This time period is called “the frame shift,” which is usually 5 to 10 ms. In this paper, frame shift is re-considered from two aspects. The first one is the appropriateness of 10 ms as the frame shift. The frame-based process is based on the assumption that temporal change of speech spectrum is slow enough compared with the frame shift, which does not hold for kinds of consonants such as plosives. Thus, this paper experimentally shows that feature value fluctuates much according to the first position of the frame. Then a training method is proposed that uses temporally shifted samples as independent samples to compensate for the fluctuation of feature caused by the difference of the beginning position of a frame. The second aspect is that the frame shift could be longer if the fluctuation can be compensated. To prove this, an experiment was conducted to change frame shift from 10 to 60 ms, and it was found that the result of 40 ms frame shift outperformed the result of 10 ms frame shift, and comparable recognition performance with 10 ms frame shift result was obtained with 50 ms frame shift.