Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation

Since the success of a time-domain speech separation, further improvements have been made by expanding the length and channel of a feature sequence to increase the amount of computation. When temporally expanded to a long sequence, the feature is segmented into chunks as a dual-path model in most studies of speech separation. In particular, it is common for the process of separating features corresponding to each speaker to be located in the final stage of the network. However, it is more advantageous and intuitive to proactively expand the feature sequence to include the number of speakers as an extra dimension. In this paper, we present an asymmetric strategy in which the encoder and decoder are partitioned to perform distinct processing in separation tasks. The encoder analyzes features, and the output of the encoder is split into the number of speakers to be separated. The separated sequences are then reconstructed by the weight-shared decoder, as Siamese network, in addition to cross-speaker processing. By using the Siamese network in the decoder, without using speaker information, the network directly learns to discriminate the features using a separation objective. With a common split layer, intermediate encoder features for skip connections are also split for the reconstruction decoder based on the U-Net structure. In addition, instead of segmenting the feature into chunks as dual-path, we design global and local Transformer blocks to directly process long sequences. The experimental results demonstrated that this separation-and-reconstruction framework is effective and that the combination of proposed global and local Transformer can sufficiently replace the role of inter- and intra-chunk processing in dual-path structure. Finally, the presented model including both of these achieved state-of-the-art performance with less computation than before in various benchmark datasets.

NeXt-TDNN: Modernizing Multi-Scale Temporal Convolution Backbone for Speaker Verification

In speaker verification, ECAPA-TDNN has shown remarkable improvement by utilizing one-dimensional(1D) Res2Net block and squeeze-and-excitation(SE) module, along with multi-layer feature aggregation (MFA). Meanwhile, in vision tasks, ConvNet structures have been modernized by referring to Transformer, resulting in improved performance. In this paper, we present an improved block design for TDNN in speaker verification. Inspired by recent ConvNet structures, we replace the SE-Res2Net block in ECAPA-TDNN with a novel 1D two-step multi-scale ConvNeXt block, which we call TS-ConvNeXt. The TS-ConvNeXt block is constructed using two separated sub-modules: a temporal multi-scale convolution (MSC) and a frame-wise feed-forward network (FFN). This two-step design allows for flexible capturing of inter-frame and intra-frame contexts. Additionally, we introduce global response normalization (GRN) for the FFN modules to enable more selective feature propagation, similar to the SE module in ECAPA-TDNN. Experimental results demonstrate that NeXt-TDNN, with a modernized backbone block, significantly improved performance in speaker verification tasks while reducing parameter size and inference time. We have released our code for future studies.

Statistical Beamformer Exploiting Non-stationarity and Sparsity with Spatially Constrained ICA for Robust Speech Recognition

In this paper, we present a statistical beamforming algorithm as a pre-processing step for robust automatic speech recognition (ASR). By modeling the target speech as a non-stationary Laplacian distribution, a mask-based statistical beamforming algorithm is proposed to exploit both its output and masked input variance for robust estimation of the beamformer. In addition, we also present a method for steering vector estimation (SVE) based on a noise power ratio obtained from the target and noise outputs in independent component analysis (ICA). To update the beamformer in the same ICA framework, we derive ICA with distortionless and null constraints on target speech, which yields beamformed speech at the target output and noises at the other outputs, respectively. The demixing weights for the target output result in a statistical beamformer with the weighted spatial covariance matrix (wSCM) using a weighting function characterized by a source model. To enhance the SVE, the strict null constraints imposed by the Lagrange multiplier methods are relaxed by generalized penalties with weight parameters, while the strict distortionless constraints are maintained. Furthermore, we derive an online algorithm based on an optimization technique of recursive least squares (RLS) for practical applications. Experimental results on various environments using CHiME-4 and LibriCSS datasets demonstrate the effectiveness of the presented algorithm compared to conventional beamforming and blind source extraction (BSE) based on ICA on both batch and online processing.