23 March 2021 in IEEE Transactions on Image Processing
Robust text reading is a very challenging problem, due to the distribution of text images changing significantly in real-world scenarios. One effective solution is to align the distribution between different domains by domain adaptation methods. However, we found that these methods might struggle when dealing sequence-like text images. An important reason is that conventional domain adaptation methods strive to align images as a whole, while text images consist of variable-length fine-grained character information. To address this issue, we propose a novel Adversarial Sequence-to-Sequence Domain Adaptation (ASSDA) method to learn “where to adapt” and “how to align” the sequential image. Our key idea is to mine the local regions that contain characters, and focus on aligning them across domains in an adversarial manner. Extensive text recognition experiments show the ASSDA could efficiently transfer sequence knowledge and validate the promising power towards the various domain shift in the real world applications.
2 July 2018 in IEEE/ACM Transactions on Audio, Speech, and Language Processing
Deep learning based speech separation usually uses a supervised algorithm to learn a mapping function from noisy features to separation targets. These separation targets, either ideal masks or magnitude spectrograms, have prominent spectro-temporal structures. Nonnegative matrix factorization (NMF) is a well-known representation learning technique that is capable of capturing the basic spectral structures. Therefore, the combination of deep learning and NMF as an organic whole is a smart strategy. However, previous methods typically use deep neural networks (DNN) and NMF for speech separation in a separate manner. In this paper, we propose a jointly combinatorial scheme to concentrate the strengths of both DNN and NMF for speech separation. NMF is used to learn the basis spectra that then are integrated into a DNN to directly reconstruct the magnitude spectrograms of speech and noise. Instead of predicting activation coefficients inferred by NMF which is used as an intermediate target by the previous methods, DNN directly optimizes an actual separation objective in our system, so that the accumulated errors could be alleviated. Moreover, we explore a discriminative training objective with sparsity constraints to suppress noise and preserve more speech components further. Systematic experiments show that the proposed models are competitive with the previous methods.
1 September 2021 in Neural Networks
16 October 2013 in The Journal of the Acoustical Society of America
In this paper, a computational goal for a monaural speech separation system is proposed. Since this goal is derived by maximizing the signal-to-noise ratio (SNR), it is called the optimal ratio mask (ORM). Under the approximate W-Disjoint Orthogonality assumption which almost always holds due to the sparse nature of speech, theoretical analysis shows that the ORM can improve the SNR about 10log102 dB over the ideal ratio mask. With three kinds of real-world interference, the speech separation results of SNR gain and objective quality evaluation demonstrate the correctness of the theoretical analysis, and imply that the ORM achieves a better separation performance.
29 September 2020 in IEEE Transactions on Circuits and Systems II: Express Briefs
This brief presents a voltage reference with high-order curvature compensation based on the zero temperature coefficient (ZTC) of MOSFET. The proposed voltage reference uses a conventional current reference to generate a constant current as of the bias of ZTCMOS. A curvature compensation method based on the α power model is combined with the conventional curvature compensation method to obtain a low temperature coefficient. Test results of the proposed voltage reference fabricated with the CSMC 0.18 μm CMOS process demonstrate that the output voltage is 628 mV. The trimmed temperature coefficient achieves 2.5 ppm/°C. The line sensitivity is 0.03 %/V, the chip area is 0.024 mm 2 , power consumption is 77 μW. The simulated power supply rejection ratio (PSRR) reaches -91.4 dB at 100 MHz.
19 July 2012 in IEEE Signal Processing Letters
In present binary masking based speech separation systems, it is almost impossible to obtain the ideal binary mask (IBM). The error in IBM estimation usually results in energy absence in many speech-dominated time-frequency (T-F) units. It violates smooth evolution nature of the speech signal and creates great artefacts. Markov random field (MRF) is one of the promising approaches to model smooth evolution nature which has been extensively applied to image smoothing applications. In this letter, an MRF prior for modeling the spatial dependencies in audio cochleagram is introduced. With this prior model, we further smooth the binary mask based cochleagram and generalize binary mask to ratio mask via a Bayesian framework. Our algorithm is systematically evaluated and compared with other counterpart methods, and it yields substantially better performance, especially on suppressing artefacts.