A field that has directly benefited from the recent advances in deep learning is automatic speech recognition (ASR). Despite the great achievements of the past decades, however, a natural and robust human–machine speech interaction still appears to be out of reach, especially in challenging environments characterized by significant noise and reverberation. To improve robustness, modern speech recognizers often employ acoustic models based on recurrent neural networks (RNNs) that are naturally able to exploit large time contexts and long-term speech modulations. It is thus of great interest to continue the study of proper techniques for improving the effectiveness of RNNs in processing speech signals. In this paper, we revise one of the most popular RNN models, namely, gated recurrent units (GRUs), and propose a simplified architecture that turned out to be very effective for ASR. The contribution of this work is twofold: First, we analyze the role played by the reset gate, showing that a significant redundancy with the update gate occurs. As a result, we propose to remove the former from the GRU design, leading to a more efficient and compact single-gate model. Second, we propose to replace hyperbolic tangent with rectified linear unit activations. This variation couples well with batch normalization and could help the model learn long-term dependencies without numerical issues. Results show that the proposed architecture, called light GRU, not only reduces the per-epoch training time by more than 30% over a standard GRU, but also consistently improves the recognition accuracy across different tasks, input features, noisy conditions, as well as across different ASR paradigms, ranging from standard DNN-HMM speech recognizers to end-to-end connectionist temporal classification models.

This paper proposes a new method of frequency-domain blind source separation (FD-BSS), able to separate acoustic sources in challenging conditions. In frequency-domain BSS, the time-domain signals are transformed into time-frequency series and the separation is generally performed by applying independent component analysis (ICA) at each frequency envelope. When short signals are observed and long demixing filters are required, the number of time observations for each frequency is limited and the variance of the ICA estimator increases due to the intrinsic statistical bias. Furthermore, common methods used to solve the permutation problem fail, especially with sources recorded under highly reverberant conditions. We propose a recursively regularized implementation of the ICA (RR-ICA) that overcomes the mentioned problem by exploiting two types of deterministic knowledge: 1) continuity of the demixing matrix across frequencies; 2) continuity of the time-activity of the sources. The recursive regularization propagates the statistics of the sources across frequencies reducing the effect of statistical bias and the occurrence of permutations. Experimental results on real-data show that the algorithm can successfully perform a fast separation of short signals (e.g., 0.5-1s), by estimating long demixing filters to deal with highly reverberant environments (e.g., ms).