A study on methods to reduce sinusoidal noises superposing on speech signal

摘要

This report discuss about the reduction of a sinusoidal noise which was difficult in the conventional technique which reduces the noise. Since a sinusoidal noise will have the characteristic just like the vowel, distinction between a speech and noise is difficult in the conventional technique. so reduction of a sinusoidal noise was difficult. Then, its attention is paid to a slightly different portion on the characteristic of a sound and a sinusoidal noise and the noise reduction is realized by an Linear Prediction Filter included a Delay Circuit. That is, the Linear Prediction Filter, conventional Linear Prediction Filter ask for the amount of data about future time t + τ (τ > 0), ask for the amount of data about future time t + τ (τ' = τ + delay) using the observation data to time t. So, it is expected that only a sinusoidal noise out from the Linear Prediction Filter by it, because it is predicted that the correlation of a sinusoidal noise is maintained although the correlation of a speech fades. Therefore, if the difference (prediction error) of the purpose signal and the output of the Linear Prediction Filter is taken, the signal with which sinusoidal noise was reduced can be acquired. Furthermore, this report discuss the technique which prefaces a Band Pass Filter to the above-mentioned technique.