Online Direction Of Arrival Estimation Based On Deep Learning

As one of the most common communication media in our lives,speech is getting more and more attention.With the rapid development of artificial intelligence(AI)in recent years,speech interaction becomes more and more important.Microphone array signal processing plays an import role in speech interaction.As an important research topic in the microphone array signal processing,direction of arrival(DOA)has been widely used in robotics,smart home and other areas.Accurate and efficient DOA estimation directly affects the efficiency of speech interaction,smart products for the experience and so on.DOA estimation is relative easy in a clean environment.However,in the natural conditions,the main challenges of DOA estimation are the interference of background noise and room reverberation on the source sound.Although this problem has been studied for decades,it is still very challenging.In this paper,we proposed a method that combines the convolutional neural network(CNN)and long short term memory(LSTM)for accurate DOA estimation in noisy and heavy reverberation environments.We evaluate the proposed method and compare it with different methods on the same dataset.The results show that the proposed method outperforms the other methods in accuracy rate(AR)insignificantly.Compared with the conventional method such as time difference of arrival(TDOA)based on generalized cross correlation(GCC)and deep learning based method,AR of the proposed method improves 41.3% and 31.5% on average,voice decision error(VDE)of the proposed method drops 20.6% on average compared with CNN.More than this,we evaluate the method on the unmatched test set.Experimental results show that the proposed method has a good robustness for the microphone array topology.The trained model can adapt to a new microphone array with a few data.