Speech Separation in Time–frequency Domain by Deep Learning with High Performance and Reducing Parameters

Abstract Introduction Traditional speech separation algorithms used time–frequency (T-F) masks. Recently, deep neural networks, which are generally called deep learning, have been used to estimate T-F masks, and the separation performance has been improved owing to it. However, a conventional deep learning approach that achieved the high separation performance requires a large number of parameters because its network is composed of recurrent neural networks (RNNs) which have recurrent and full connections. This is a disadvantage because the low memory consumption is desired in the case of using speech separation algorithms for actual applications. Because of this shortcoming, we proposed a novel network architecture that balances both the high separation performance and the low memory cost. Methods The proposed network is composed of convolutional neural networks (CNNs) which have a sparse connection instead of RNNs to reduce the number of parameters. Additionally, aiming to achieve the high separation performance, we designed the proposed network for speech separation to make it possible to learn the feature of speech. We evaluated the separation performance for each network on the task of separating two-speaker mixture. Results In the simulation results, the separation performance of the proposed network was competitive with one of conventional ones with the highest separation performance regardless of realizing the more than 80% reduction of the number of parameters. Conclusion The results indicated that the proposed network is superior to a conventional one considering the cost-performances of both the speech separation performance and the memory cost.