This paper presents a joint learning algorithm for complex-valued time-frequency (T-F) masks in single-channel speech enhancement systems. Most speech enhancement algorithms operating in a single-channel microphone environment aim to enhance the magnitude component in a T-F domain, while the input noisy phase component is used directly without any processing. Consequently, the mismatch between the processed magnitude and the unprocessed phase degrades the sound quality. To address this issue, a learning method of targeting a T-F mask that is defined in a complex domain has recently been proposed. However, due to a wide dynamic range and an irregular spectrogram pattern of the complex-valued T-F mask, the learning process is difficult even with a large-scale deep learning network. Moreover, the learning process targeting the T-F mask itself does not directly minimize the distortion in spectra or time domains. In order to address these concerns, we focus on three issues: 1) an effective estimation of complex numbers with a wide dynamic range; 2) a learning method that is directly related to speech enhancement performance; and 3) a way to resolve the mismatch between the estimated magnitude and phase spectra. In this study, we propose objective functions that can solve each of these issues and train the network by minimizing them with a joint learning framework. The evaluation results demonstrate that the proposed learning algorithm achieves significant performance improvement in various objective measures and subjective preference listening test.

This letter presents a phase-sensitive joint learning algorithm for single-channel speech enhancement. Although a deep learning framework that estimates the time-frequency (T-F) domain ideal ratio masks demonstrates a strong performance, it is limited in the sense that the enhancement process is performed only in the magnitude domain, while the phase spectra remain unchanged. Thus, recent studies have been conducted to involve phase spectra in speech enhancement systems. A phase-sensitive mask (PSM) is a T-F mask that implicitly represents phase-related information. However, since the PSM has an unbounded value, the networks are trained to target its truncated values rather than directly estimating it. To effectively train the PSM, we first approximate it to have a bounded dynamic range under the assumption that speech and noise are uncorrelated. We then propose a joint learning algorithm that trains the approximated value through its parameterized variables in order to minimize the inevitable error caused by the truncation process. Specifically, we design a network that explicitly targets three parameterized variables: 1) speech magnitude spectra; 2) noise magnitude spectra; and 3) phase difference of clean to noisy spectra. To further improve the performance, we also investigate how the dynamic range of magnitude spectra controlled by a warping function affects the final performance in joint learning algorithms. Finally, we examined how the proposed additional constraint that preserves the sum of the estimated speech and noise power spectra affects the overall system performance. The experimental results show that the proposed learning algorithm outperforms the conventional learning algorithm with the truncated phase-sensitive approximation.

This paper presents a speech emotion recognition system using a recurrent neural network (RNN) model trained by an efficient learning algorithm. The proposed system takes into account the long-range context effect and the uncertainty of emotional label expressions. To extract high-level representation of emotional states with regard to its temporal dynamics, a powerful learning method with a bidirectional long short-term memory (BLSTM) model is adopted. To overcome the uncertainty of emotional labels, such that all frames in the same utterance are mapped into the same emotional label, it is assumed that the label of each frame is regarded as a sequence of random variables. Then, the sequences are trained by the proposed learning algorithm. The weighted accuracy of the proposed emotion recognition system is improved up to 12% compared to the DNN-ELM based emotion recognition system used as a baseline.