Recent machine learning based approaches to speech enhancement operate in the time domain and have been shown to outperform the classical enhancement methods. Two such models are SE-GAN and SE-WaveNet, both of which rely on complex neural network architectures, making them expensive to train. We propose using the Variance Constrained Autoencoder (VCAE) for speech enhancement. Our model uses a more straightforward neural network structure than competing solutions and is a natural model for the task of speech enhancement. We demonstrate experimentally that the proposed enhancement model outperforms SE-GAN and SE-WaveNet in terms of perceptual quality of enhanced signals. Index Terms: speech enhancement, generative modelling, time domain, deep learning, neural networks ditionally minimise an L1 error between the enhanced and desired signals. Since SE-VCAE optimises an L1 measure, it is also not a fully generative enhancement approach. Key contributions: 1) We propose SE-VCAE (Section 3), a novel machine learning speech enhancement approach based on VCAE (introduced in Section 2), 2) our proposed system outperforms SE-GAN and SE-WaveNet (Section 4), and 3) SE-VCAEs implementation is more straightforward than both competing methods, demonstrating that complex neural network structures are not necessary to enhance speech.

Understanding the microbial community structure and genetic potential of anode biofilms is key to improve extracellular electron transfers in microbial fuel cells. We investigated effect of substrate and temporal dynamics of anodic biofilm communities using phylogenetic and metagenomic approaches in parallel with electrochemical characterizations. The startup non-steady state anodic bacterial structures were compared for a simple substrate, acetate, and for a complex substrate, landfill leachate, using a single-chamber air-cathode microbial fuel cell. Principal coordinate analysis showed that distinct community structures were formed with each substrate type. The bacterial diversity measured as Shannon index decreased with time in acetate cycles, and was restored with the introduction of leachate. The change of diversity was accompanied by an opposite trend in the relative abundance of Geobacter-affiliated phylotypes, which were acclimated to over 40% of total Bacteria at the end of acetate-fed conditions then declined in the leachate cycles. The transition from acetate to leachate caused a decrease in output power density from 243±13 mW/m2 to 140±11 mW/m2, accompanied by a decrease in Coulombic electron recovery from 18±3% to 9±3%. The leachate cycles selected protein-degrading phylotypes within phylum Synergistetes. Metagenomic shotgun sequencing showed that leachate-fed communities had higher cell motility genes including bacterial chemotaxis and flagellar assembly, and increased gene abundance related to metal resistance, antibiotic resistance, and quorum sensing. These differentially represented genes suggested an altered anodic biofilm community in response to additional substrates and stress from the complex landfill leachate.

No abstract

Recent state-of-the-art autoencoder based generative models have an encoder-decoder structure and learn a latent representation with a pre-defined distribution that can be sampled from. Implementing the encoder networks of these models in a stochastic manner provides a natural and common approach to avoid overfitting and enforce a smooth decoder function. However, we show that for stochastic encoders, simultaneously attempting to enforce a distribution constraint and minimising an output distortion leads to a reduction in generative and reconstruction quality. In addition, attempting to enforce a latent distribution constraint is not reasonable when performing disentanglement. Hence, we propose the variance-constrained autoencoder (VCAE), which only enforces a variance constraint on the latent distribution. Our experiments show that VCAE improves upon Wasserstein Autoencoder and the Variational Autoencoder in both reconstruction and generative quality on MNIST and CelebA. Moreover, we show that VCAE equipped with a total correlation penalty term performs equivalently to FactorVAE at learning disentangled representations on 3D-Shapes while being a more principled approach.

For hospitalized patients requiring intravenous insulin therapy, an objective is to quantify the intravenous insulin infusion rate (IR) across the domain of blood glucose (BG) values at a single timepoint. The algorithm parameters include low BG (70 mg/dL), critical high BG, target range BG limits, and maintenance rate (MR) of insulin infusion, which, after initialization, depends on rate of change of blood glucose, previous IR, and other inputs. The restraining rate (RR) is a function of fractional completeness of ascent of BG (FCABG) from BG 70 mg/dL to target. The correction rate (CR) is a function of fractional elevation of BG (FEBG), in comparison to elevation of a critical high BG, above target. IR = RR + CR. The proposed mathematical model describing a sigmoidal relationship between IR and BG may offer a safety advantage over the linear relationship currently employed in some intravenous glucose management systems.