類神經網路已能在聲音辨識上取得極好的成績，多種不同的聲音特徵都被嘗試作為網路的輸入進行訓練辨識，然而以原始聲音訊號作為網路輸入，測試神經網路是否能夠自行擷取出聲音特徵依舊是一門挑戰。本文改良了現有原始訊號網路的架構，利用高層數的深度神經網路成功提升了訊號輸入分析的效果，以擬似頻譜轉換的方式，探討正確的參數設定，最終提出的1d-2d network 於ESC50中可成功達到73.55%的正確率。 除此之外，本文亦提出一種特徵融合的網路架構，利用全域池化層的特性，整合出一種較具彈性的結合方式。利用此類網路，本文成功結合了利用原始訊號輸入及利用對數梅爾頻譜係數輸入的兩種網路，我們提出的ParallelNet在ESC50中以上述輸入得到了81.55%的辨識效果，達到了人類辨識水平。 ;Neural network has achieved a great result in the sound recognition, many different kinds of acoustic features have been tried as the training input with the network. However, there is still under doubt that the whether the neural network could efficiently extract features from the raw audio signal input. This study improved the raw-signal-input network from other researches, with the deeper network architectures, the raw signals get the well analysis with our network, we also make the discussion in several kinds of network settings, with the spectrogram-like conversion, our network could reach the accuracy of 73.55% in the open-audio-dataset ESC50. Besides, in this study, we proposed a network architectures that could combine different kinds of networks feed with different features. With the help of global pooling, a flexible fusion way is well integrated into the network. Our experiment successfully combined two different networks which use different kinds of audio feature inputs—raw audio signal and log-mel spectrum. By the above settings, the ParallelNet we proposed finally reaches the accuracy of 81.55% in ESC50, which also reaches the recognition level of human being.
