The complexity of polyphonic sounds imposes numerous challenges on their classification. Especially in real life, polyphonic sound events have discontinuity and unstable time-frequency variations. Traditional single acoustic features cannot characterize the key feature information of the polyphonic sound event, and this deficiency results in poor model classification performance. In this paper, we propose a convolutional recurrent neural network model based on the temporal-frequency (TF) attention mechanism and feature space (FS) attention mechanism (TFFS-CRNN). The TFFS-CRNN model aggregates Log-Mel spectrograms and MFCCs feature as inputs, which contains the TF-attention module, the convolutional recurrent neural network (CRNN) module, the FS-attention module and the bidirectional gated recurrent unit (BGRU) module. In polyphonic sound events detection (SED), the TF-attention module can capture the critical temporal–frequency features more capably. The FS-attention module assigns different dynamically learnable weights to different dimensions of features. The TFFS-CRNN model improves the characterization of features for key feature information in polyphonic SED. By using two attention modules, the model can focus on semantically relevant time frames, key frequency bands, and important feature spaces. Finally, the BGRU module learns contextual information. The experiments were conducted on the DCASE 2016 Task3 dataset and the DCASE 2017 Task3 dataset. Experimental results show that the F1-score of the TFFS-CRNN model improved 12.4% and 25.2% compared with winning system models in DCASE challenge; the ER is reduced by 0.41 and 0.37 as well. The proposed TFFS-CRNN model algorithm has better classification performance and lower ER in polyphonic SED.

Sensor arrays are gradually becoming a current research hotspot due to their flexible beam control, high signal gain, robustness against extreme interference, and high spatial resolution. Among them, spherical microphone arrays with complex rotational symmetry can capture more sound field information than planar arrays and can convert the collected multiple speech signals into the spherical harmonic domain for processing through spherical modal decomposition. The subspace class direction of arrival (DOA) estimation algorithm is sensitive to noise and reverberation, and its performance can be improved by introducing relative sound pressure and frequency-smoothing techniques. The introduction of the relative sound pressure can increase the difference between the eigenvalues corresponding to the signal subspace and the noise subspace, which is helpful to estimate the number of active sound sources. The eigenbeam estimation of signal parameters via the rotational invariance technique (EB-ESPRIT) is a well-known subspace-based algorithm for a spherical microphone array. The EB-ESPRIT cannot estimate the DOA when the elevation angle approaches 90°. Huang et al. proposed a two-step ESPRIT (TS-ESPRIT) algorithm to solve this problem. The TS-ESPRIT algorithm estimates the elevation and azimuth angles of the signal independently, so there is a problem with DOA parameter pairing. In this paper, the DOA parameter pairing problem of the TS-ESPRIT algorithm is solved by introducing generalized eigenvalue decomposition without increasing the computation of the algorithm. At the same time, the estimation of the elevation angle is given by the arctan function, which increases the estimation accuracy of the elevation angle of the algorithm. The robustness of the algorithm in a noisy environment is also enhanced by introducing the relative sound pressure into the algorithm. Finally, the simulation and field-testing results show that the proposed method not only solves the problem of DOA parameter pairing, but also outperforms the traditional methods in DOA estimation accuracy.