Search for: All records

In this study, we propose to investigate triplet loss for the purpose of an alternative feature representation for ASR. We consider a general non-semantic speech representation, which is trained with a self-supervised criteria based on triplet loss called TRILL, for acoustic modeling to represent the acoustic characteristics of each audio. This strategy is then applied to the CHiME-4 corpus and CRSS-UTDallas Fearless Steps Corpus, with emphasis on the 100-hour challenge corpus which consists of 5 selected NASA Apollo-11 channels. An analysis of the extracted embeddings provides the foundation needed to characterize training utterances into distinct groups based on acoustic distinguishing properties. Moreover, we also demonstrate that triplet-loss based embedding performs better than i-Vector in acoustic modeling, confirming that the triplet loss is more effective than a speaker feature. With additional techniques such as pronunciation and silence probability modeling, plus multi-style training, we achieve a +5.42% and +3.18% relative WER improvement for the development and evaluation sets of the Fearless Steps Corpus. To explore generalization, we further test the same technique on the 1 channel track of CHiME-4 and observe a +11.90% relative WER improvement for real test data. 

The climatic feedbacks from vegetation, particularly from tropical forests, can alter climate through land‐atmospheric interactions. Expected shifts in species composition can alter these interactions with profound effects on climate and terrestrial ecosystem dynamics. Ecosystem demographic (ED) models can explicitly represent vegetation dynamics and are a key component of next‐generation Earth System Models (ESMs). Although ED models exhibit greater fidelity and allow more direct comparisons with observations, their interacting parameters can be more difficult to calibrate due to the complex interactions among vegetation groups and physical processes. In addition, while representation of forest successional coexistence in ESMs is necessary to accurately capture forest‐climate interactions, few models can simulate forest coexistence and few studies have calibrated coexisted forest species. Furthermore, although both vegetation characteristics and soil properties affect vegetation dynamics, few studies have paid attention to jointly calibrating parameters related to these two processes. In this study, we develop a computationally‐efficient and physical model structure‐based framework that uses a parallel surrogate global optimization algorithm to calibrate ED models. We calibrate two typically coexisted tropical tree species, early and late successional plants, in a state‐of‐the‐art ED model that is capable of simulating successional diversity in forests. We concurrently calibrate vegetation and soil parameters and validate results against carbon, energy, and water cycle measurements collected in Barro Colorado Island, Panama. The framework can find optimal solutions within 4–12 iterations for 19‐dimensional problems. The calibration for tropical forests has important implications for predicting land‐atmospheric interactions and responses of tropical forests to environmental changes.

 

Materials with giant spin splitting are desired for spintronic applications. The fabrications of spintronic devices from half metals with one spin direction are often hampered, however, by stray magnetic fields, domain walls, short spin coherence times, scattering on magnetic atoms or magnetically active interfaces, and other characteristics that come along with the magnetism. The surfaces of topological insulators, or Dirac/Weyl semimetals, could be an alternative, but production of high‐quality thin films without the presence of the bulk states at the Fermi energy remains very challenging. Here, by utilizing angle‐resolved photoemission spectroscopy, a record‐high Dresselhaus spin–orbit splitting of the bulk state in the nonmagnetic IrBiSe is found. The band structure calculations indicate that the splitting band is fully spin‐polarized with 3D chiral spin texture. As a source of spin‐polarized electrons, lightly doped IrBiSe is expected to generate electric‐field‐controlled spin‐polarized currents, free from back scattering, and could host triplet and Fulde–Ferrel–Larkin–Ovchinnikov (FFLO) superconductivity.