Search for: All records

Neural codecs have demonstrated strong performance in high-fidelity compression of audio signals at low bitrates. The token-based representations produced by these codecs have proven particularly useful for generative modeling. While much research has focused on improvements in compression ratio and perceptual transparency, recent works have largely overlooked another desirable codec property -- \textit{idempotence}, the stability of compressed outputs under multiple rounds of encoding. We find that state-of-the-art neural codecs exhibit varied degrees of idempotence, with some degrading audio outputs significantly after as few as three encodings. We investigate possible causes of low idempotence and devise a method for improving idempotence through fine-tuning a codec model. We then examine the effect of idempotence on a simple conditional generative modeling task, and find that increased idempotence can be achieved without negatively impacting downstream modeling performance -- potentially extending the usefulness of neural codecs for practical file compression and iterative generative modeling workflows. 

The destructive 2023 moment magnitude ( M w ) 7.8-7.7 earthquake doublet ruptured multiple segments of the East Anatolian Fault system in Turkey. We integrate multi-scale seismic and space-geodetic observations with multi-fault kinematic inversions and dynamic rupture modeling to unravel the events’ complex rupture history and stress-mediated fault interactions. Our analysis reveals three sub-shear slip episodes during the initial M w 7.8 earthquake with delayed rupture initiation to the southwest. The M w 7.7 event occurred 9 hours later with larger slip and supershear rupture on its western branch. Mechanically consistent dynamic models accounting for fault interactions can explain the unexpected rupture paths, and require a heterogeneous background stress. Our results highlight the importance of combining near- and far-field observations with data-driven and physics-based models for seismic hazard assessment. 

ABSTRACT The July 2019 Ridgecrest, California, earthquake sequence involved two large events—the M 6.4 foreshock and the M 7.1 mainshock that ruptured a system of intersecting strike-slip faults. We present analysis of space geodetic observations including Synthetic Aperture Radar and Global Navigation Satellite System data, geological field mapping, and seismicity to constrain the subsurface rupture geometry and slip distribution. The data render a complex pattern of faulting with a number of subparallel as well as cross-cutting fault strands that exhibit variations in both strike and dip angles, including a “flower structure” formed by shallow splay faults. Slip inversions are performed using both homogeneous and layered elastic half-space models informed by the local seismic tomography data. The inferred slip distribution suggests a moderate amount of the shallow coseismic slip deficit. The peak moment release occurred in the depth interval of 3–4 km, consistent with results from previous studies of major strike-slip earthquakes, and the depth distribution of seismicity in California. We use the derived slip models to investigate stress transfer and possible triggering relationships between the M 7.1 mainshock and the M 6.4 foreshock, as well as other moderate events that occurred in the vicinity of the M 7.1 hypocenter. Triggering is discouraged for the average strike of the M 7.1 rupture (320°) but encouraged for the initial orientation of the mainshock rupture suggested by the first-motion data (340°). This lends support to a scenario according to which the earthquake rupture nucleated on a small fault that was more optimally oriented with respect to the regional stress and subsequently propagated along the less-favorably oriented pre-existing faults, possibly facilitated by dynamic weakening. The nucleation site of the mainshock experienced positive dynamic Coulomb stress changes that are much larger than the static stress changes, yet the former failed to initiate rupture. 

Non-parallel many-to-many voice conversion remains an interesting but challenging speech processing task. Many style-transfer-inspired methods such as generative adversarial networks (GANs) and variational autoencoders (VAEs) have been proposed. Recently, AUTOVC, a conditional autoencoders (CAEs) based method achieved state-of-the-art results by disentangling the speaker identity and speech content using information-constraining bottlenecks, and it achieves zero-shot conversion by swapping in a different speaker’s identity embedding to synthesize a new voice. However, we found that while speaker identity is disentangled from speech content, a significant amount of prosodic information, such as source F0, leaks through the bottleneck, causing target F0 to fluctuate unnaturally. Furthermore, AUTOVC has no control of the converted F0 and thus unsuitable for many applications. In the paper, we modified and improved autoencoder-based voice conversion to disentangle content, F0, and speaker identity at the same time. Therefore, we can control the F0 contour, generate speech with F0 consistent with the target speaker, and significantly improve quality and similarity. We support our improvement through quantitative and qualitative analysis. 

Abstract The 2021 Maduo earthquake ruptured a 150 km‐long left‐lateral fault in the northeast Tibet. We used Synthetic Aperture Radar data collected by the Sentinel‐1A/B satellites within days of the earthquake to derive a finite fault model and investigate the details of slip distribution with depth. We generated coseismic interferograms and pixel offsets from different look directions corresponding to the ascending and descending satellite orbits. At the eastern end the rupture bifurcated into two sub‐parallel strands, with larger slip on the northern strand. Inversions of coseismic displacements show maximum slip to the east of the epicenter. The averaged coseismic slip has a peak at depth of 3–4 km, similar to slip distributions of a number of shallow strike‐slip earthquakes. Postseismic observations over several weeks following the Maduo earthquake reveal surface slip with amplitude up to 0.1 m that at least partially eliminated the coseismic slip deficit in the uppermost crust. 

Significance Light-driven rhodopsin proteins pump ions across cell membranes. They have applications in optogenetics and can potentially be used to develop solar energy–harvesting devices. A detailed understanding of rhodopsin dynamics and functions may therefore assist research in medicine, health, and clean energy. This time-resolved crystallography study carried out with X-ray free-electron lasers reveals detailed dynamics of chloride ion–pumping rhodopsin (ClR) within 100 ps of light activation. It shows the dissociation of Cl⁻from the Schiff base binding site upon light-triggered retinal isomerization. This Cl⁻dissociation is followed by diffusion toward the intracellular direction. The results hint at a common ion-pumping mechanism across rhodopsin families.