聲音在日常生活中扮演著重要的地位，但大多環境內往往會有殘響的存在，例如視訊會議、遠距教學甚至是手機通訊等面對的議題，因此語音的清晰度顯得格外為重要。 深層類神經網路(Deep Neural Network, DNN)目前已經成為處理訊號問題的熱門方法。本論文主要以深層網路為基礎設計一個不同於以往的新架構，結合了自編碼器與深層遞迴類神經網路，稱之序列至序列自編碼模型(Sequence to sequence Autoencoder, SA)，作法是用經由短時傅立葉轉換後，將能量資訊(magnitude)輸入至網路模型，藉由同時考慮能量的時間關係和自身的結構資訊，輸出為預估的能量大小，並結合相位資訊(phase)映射回時域上。最後，本論文提出的方法使用Chime4和REVERB challenge 2014的資料作評估，實驗結果顯示本方法較其他深度類神經網路更加優秀。 ;Sound plays an important role in daily life, but most of the environment often has reverberations, such as video conferencing, distance education, and even mobile communication. Therefore, the clarity of speech is particularly important. The Deep Neural Network (DNN) has become a popular method for dealing with signal problems. This paper mainly designs a new architecture different from the previous one based on the deep network. It combines the Auto-Encoder and deep recursive neural network, called the sequence to sequence Autoencoder (SA). The method is to input the magnitude into the network model by using the energy of output of the short-time Fourier transform. Considering the temporal relationship of energy and its structural information, the estimated energy is output and then combined with the phase information to map to the time domain. Finally, the proposed method in this paper uses Chime4 and REVERB challenge 2014 data for reverberation elimination. The experimental results show that this method is superior than other deep neural networks.
