Speech Enhancement using Separable Polling Attention and Global Layer Normalization followed with PReLU

Single channel speech enhancement is a challenging task in speech community. Recently, various neural networks based methods have been applied to speech enhancement. Among these models, PHASEN and T-GSA achieve state-of-the-art performances on the publicly opened VoiceBank+DEMAND corpus. Both of the models reach the COVL score of 3.62. PHASEN achieves the highest CSIG score of 4.21 while T-GSA gets the highest PESQ score of 3.06. However, both of these two models are very large. The contradiction between the model performance and the model size is hard to reconcile. In this paper, we introduce three kinds of techniques to shrink the PHASEN model and improve the performance. Firstly, seperable polling attention is proposed to replace the frequency transformation blocks in PHASEN. Secondly, global layer normalization followed with PReLU is used to replace batch normalization followed with ReLU. Finally, BLSTM in PHASEN is replaced with Conv2d operation and the phase stream is simplified. With all these modifications, the size of the PHASEN model is shrunk from 33M parameters to 5M parameters, while the performance on VoiceBank+DEMAND is improved to the CSIG score of 4.30, the PESQ score of 3.07 and the COVL score of 3.73.

A multi-view approach for Mandarin non-native mispronunciation verification

Traditionally, the performance of non-native mispronunciation verification systems relied on effective phone-level labelling of non-native corpora. In this study, a multi-view approach is proposed to incorporate discriminative feature representations which requires less annotation for non-native mispronunciation verification of Mandarin. Here, models are jointly learned to embed acoustic sequence and multi-source information for speech attributes and bottleneck features. Bidirectional LSTM embedding models with contrastive losses are used to map acoustic sequences and multi-source information into fixed-dimensional embeddings. The distance between acoustic embeddings is taken as the similarity between phones. Accordingly, examples of mispronounced phones are expected to have a small similarity score with their canonical pronunciations. The approach shows improvement over GOP-based approach by +11.23% and single-view approach by +1.47% in diagnostic accuracy for a mispronunciation verification task.

Convolution Forgetting Curve Model for Repeated Learning

Most of mathematic forgetting curve models fit well with the forgetting data under the learning condition of one time rather than repeated. In the paper, a convolution model of forgetting curve is proposed to simulate the memory process during learning. In this model, the memory ability (i.e. the central procedure in the working memory model) and learning material (i.e. the input in the working memory model) is regarded as the system function and the input function, respectively. The status of forgetting (i.e. the output in the working memory model) is regarded as output function or the convolution result of the memory ability and learning material. The model is applied to simulate the forgetting curves in different situations. The results show that the model is able to simulate the forgetting curves not only in one time learning condition but also in multi-times condition. The model is further verified in the experiments of Mandarin tone learning for Japanese learners. And the predicted curve fits well on the test points.

Automated deconvolution of structured mixtures from bulk tumor genomic data

Motivation: As cancer researchers have come to appreciate the importance of intratumor heterogeneity, much attention has focused on the challenges of accurately profiling heterogeneity in individual patients. Experimental technologies for directly profiling genomes of single cells are rapidly improving, but they are still impractical for large-scale sampling. Bulk genomic assays remain the standard for population-scale studies, but conflate the influences of mixtures of genetically distinct tumor, stromal, and infiltrating immune cells. Many computational approaches have been developed to deconvolute these mixed samples and reconstruct the genomics of genetically homogeneous clonal subpopulations. All such methods, however, are limited to reconstructing only coarse approximations to a few major subpopulations. In prior work, we showed that one can improve deconvolution of genomic data by leveraging substructure in cellular mixtures through a strategy called simplicial complex inference. This strategy, however, is also limited by the difficulty of inferring mixture structure from sparse, noisy assays. Results: We improve on past work by introducing enhancements to automate learning of substructured genomic mixtures, with specific emphasis on genome-wide copy number variation (CNV) data. We introduce methods for dimensionality estimation to better decompose mixture model substructure; fuzzy clustering to better identify substructure in sparse, noisy data; and automated model inference methods for other key model parameters. We show that these improvements lead to more accurate inference of cell populations and mixture proportions in simulated scenarios. We further demonstrate their effectiveness in identifying mixture substructure in real tumor CNV data. Availability: Source code is available at http://www.cs.cmu.edu/~russells/software/WSCUnmix.zip

Derivative-free optimization of rate parameters of capsid assembly models from bulk in vitro data

The assembly of virus capsids from free coat proteins proceeds by a complicated cascade of association and dissociation steps, the great majority of which cannot be directly experimentally observed. This has made capsid assembly a rich field for computational models to attempt to fill the gaps in what is experimentally observable. Nonetheless, accurate simulation predictions depend on accurate models and there are substantial obstacles to model inference for such systems. Here, we describe progress in learning parameters for capsid assembly systems, particularly kinetic rate constants of coat-coat interactions, by computationally fitting simulations to experimental data. We previously developed an approach to learn rate parameters of coat-coat interactions by minimizing the deviation between real and simulated light scattering data monitoring bulk capsid assembly in vitro. This is a difficult data-fitting problem, however, because of the high computational cost of simulating assembly trajectories, the stochastic noise inherent to the models, and the limited and noisy data available for fitting. Here we show that a newer classes of methods, based on derivative-free optimization (DFO), can more quickly and precisely learn physical parameters from static light scattering data. We further explore how the advantages of the approaches might be affected by alternative data sources through simulation of a model of time-resolved mass spectrometry data, an alternative technology for monitoring bulk capsid assembly that can be expected to provide much richer data. The results show that advances in both the data and the algorithms can improve model inference, with rich data leading to high-quality fits for all methods, but DFO methods showing substantial advantages over less informative data sources better representative of the current experimental practice.

Avoid Internal Loops in Steady State Flux Space Sampling

As a widely used method in metabolic network studies, Monte-Carlo sampling in the steady state flux space is known for its flexibility and convenience of carrying out different purposes, simply by alternating constraints or objective functions, or appending post processes. Recently the concept of a non-linear constraint based on the second thermodynamic law, known as Loop Law, is challenging current sampling algorithms which will inevitably give rise to the internal loops. A generalized method is proposed here to eradicate the probability of the appearance of internal loops during sampling process. Based on Artificial Centered Hit and Run (ACHR) method, each step of the new sampling process will avoid entering loop-forming subspaces. This method has been applied on the metabolic network of Helicobacter pylori with three different objective functions: uniform sampling, optimizing biomass synthesis, optimizing biomass synthesis efficiency over resources ingested. Comparison between results from the new method and conventional ACHR method shows effective elimination of loop fluxes without affecting non-loop fluxes.

Imposition of Different Optimizing Object with Non-Linear Constraints on Flux Sampling and Elimination of Free Futile Pathways

Constraint-based modeling has been widely used on metabolic networks analysis, such as biosynthetic prediction and flux optimization. The linear constraints, like mass conservation constraint, reversibility constraint, biological capacity constraint, can be imposed on linear algorithms. However, recently a non-linear constraint based on the second thermodynamic law, known as loop law, has emerged and challenged the existing algorithms. Proven to be unfeasible with linear solutions, this non-linear constraint has been successfully imposed on the sampling process. In this place, Monte - Carlo sampling with Metropolis criterion and Simulated Annealing has been introduced to optimize the Biomass synthesis of genome scale metabolic network of Helicobacter pylori (iIT341 GSM / GPR) under mass conservation constraint, biological capacity constraint, and thermodynamic constraints including reversibility and loop law. The sampling method has also been employed to optimize a non-linear objective function, the Biomass synthetic rate, which is unified by the total income number of reducible electrons. To verify whether a sample contains internal loops, an automatic solution has been developed based on solving a set of inequalities. In addition, a new type of pathway has been proposed here, the Futile Pathway, which has three properties: 1) its mass flow could be self-balanced; 2) it has exchange reactions; 3) it is independent to the biomass synthesis. To eliminate the fluxes of the Futile Pathways in the sampling results, a linear programming based method has been suggested and the results have showed improved correlations among the reaction fluxes in the pathways related to Biomass synthesis.