-
Learning to Detect Novel and Fine-Grained Acoustic Sequences Using Pretrained Audio Representations
Authors:
Vasudha Kowtha,
Miquel Espi Marques,
Jonathan Huang,
Yichi Zhang,
Carlos Avendano
Abstract:
This work investigates pretrained audio representations for few shot Sound Event Detection. We specifically address the task of few shot detection of novel acoustic sequences, or sound events with semantically meaningful temporal structure, without assuming access to non-target audio. We develop procedures for pretraining suitable representations, and methods which transfer them to our few shot le…
▽ More
This work investigates pretrained audio representations for few shot Sound Event Detection. We specifically address the task of few shot detection of novel acoustic sequences, or sound events with semantically meaningful temporal structure, without assuming access to non-target audio. We develop procedures for pretraining suitable representations, and methods which transfer them to our few shot learning scenario. Our experiments evaluate the general purpose utility of our pretrained representations on AudioSet, and the utility of proposed few shot methods via tasks constructed from real-world acoustic sequences. Our pretrained embeddings are suitable to the proposed task, and enable multiple aspects of our few shot framework.
△ Less
Submitted 3 May, 2023;
originally announced May 2023.
-
Pre-trained Model Representations and their Robustness against Noise for Speech Emotion Analysis
Authors:
Vikramjit Mitra,
Vasudha Kowtha,
Hsiang-Yun Sherry Chien,
Erdrin Azemi,
Carlos Avendano
Abstract:
Pre-trained model representations have demonstrated state-of-the-art performance in speech recognition, natural language processing, and other applications. Speech models, such as Bidirectional Encoder Representations from Transformers (BERT) and Hidden units BERT (HuBERT), have enabled generating lexical and acoustic representations to benefit speech recognition applications. We investigated the…
▽ More
Pre-trained model representations have demonstrated state-of-the-art performance in speech recognition, natural language processing, and other applications. Speech models, such as Bidirectional Encoder Representations from Transformers (BERT) and Hidden units BERT (HuBERT), have enabled generating lexical and acoustic representations to benefit speech recognition applications. We investigated the use of pre-trained model representations for estimating dimensional emotions, such as activation, valence, and dominance, from speech. We observed that while valence may rely heavily on lexical representations, activation and dominance rely mostly on acoustic information. In this work, we used multi-modal fusion representations from pre-trained models to generate state-of-the-art speech emotion estimation, and we showed a 100% and 30% relative improvement in concordance correlation coefficient (CCC) on valence estimation compared to standard acoustic and lexical baselines. Finally, we investigated the robustness of pre-trained model representations against noise and reverberation degradation and noticed that lexical and acoustic representations are impacted differently. We discovered that lexical representations are more robust to distortions compared to acoustic representations, and demonstrated that knowledge distillation from a multi-modal model helps to improve the noise-robustness of acoustic-based models.
△ Less
Submitted 3 March, 2023;
originally announced March 2023.
-
Infrared variability of young solar analogs in the Lagoon Nebula
Authors:
C. Ordenes-Huanca,
M. Zoccali,
A. Bayo,
J. Cuadra,
R. Contreras Ramos,
L. A. Hillenbrand,
I. Lacerna,
S. Abarzua,
C. Avendaño,
P. Diaz,
I. Fernandez,
G. Lara
Abstract:
T Tauri stars are low-mass pre-main sequence stars that are intrinsically variable. Due to the intense magnetic fields they possess, they develop dark spots on their surface that, because of rotation, introduce a periodic variation of brightness.In addition, the presence of surrounding disks could generate flux variations by variable extinction or accretion. Both can lead to a brightness decrease…
▽ More
T Tauri stars are low-mass pre-main sequence stars that are intrinsically variable. Due to the intense magnetic fields they possess, they develop dark spots on their surface that, because of rotation, introduce a periodic variation of brightness.In addition, the presence of surrounding disks could generate flux variations by variable extinction or accretion. Both can lead to a brightness decrease or increase, respectively. Here, we have compiled a catalog of light curves for 379 T Tauri stars in the Lagoon Nebula (M8) region, using VVVX survey data in the Ks-band. All these stars were already classified as pre-MS stars based on other indicators. The data presented here are spread over a period of about eight years, which gives us a unique follow-up time for these sources at this wavelength. The light curves were classified according to their degree of periodicity and asymmetry, to constrain the physical processes responsible for their variation. Periods were compared with the ones found in literature, on a much shorter baseline. This allowed us to prove that for 126 stars, the magnetically active regions remain stable for several years. Besides, our near-IR data were compared with the optical Kepler/K2 light curves, when available, giving us a better understanding of the mechanisms responsible for the brightness variations observed and how they manifest at different bands. We found that the periodicity in both bands is in fairly good agreement, but the asymmetry will depend on the amplitude of the bursts or dips events and the observation cadence.
△ Less
Submitted 17 October, 2022;
originally announced October 2022.
-
Empty smectics of hard nanorings: insights from a second-virial theory
Authors:
H. H. Wensink,
C. Avendaño
Abstract:
Inspired by recent simulations on highly open liquid crystalline structures formed by rigid planar nanorings we present a simple theoretical framework explaining the prevalence of smectic over nematic ordering in systems of ring-shaped objects. The key part of our study is a calculation of the excluded volume of such non-convex particles in the limit of vanishing thickness to diameter ratio. Using…
▽ More
Inspired by recent simulations on highly open liquid crystalline structures formed by rigid planar nanorings we present a simple theoretical framework explaining the prevalence of smectic over nematic ordering in systems of ring-shaped objects. The key part of our study is a calculation of the excluded volume of such non-convex particles in the limit of vanishing thickness to diameter ratio. Using a simple stability analysis we then show that dilute systems of ring-shaped particles have a strong propensity to order into smectic structures with an unusual antinematic order while solid disks of the same dimensions exhibit nematic order. Since our model rings have zero internal volume these smectic structures are essential empty, resembling the strongly porous structures found in simulation. We argue that the antinematic intralamellar order of the rings plays an essential role in stabilizing these novel smectic structures.
△ Less
Submitted 21 September, 2016;
originally announced September 2016.
-
Orientational ordering and phase behaviour of a binary mixture of hard spheres and hard spherocylinders
Authors:
Liang Wu,
Alexandr Malijevský,
George Jackson,
Erich A. Müller,
Carlos Avendaño
Abstract:
We study structure and fluid-phase behaviour of a binary mixture of hard spheres (HSs) and hard spherocylinders (HSCs) in isotropic and nematic states using the $NP_nAT$ ensemble Monte Carlo (MC) method in which a normal pressure tensor component is fixed in a system confined between two hard walls. The method allows one to estimate the location of the isotropic-nematic phase transition and to obs…
▽ More
We study structure and fluid-phase behaviour of a binary mixture of hard spheres (HSs) and hard spherocylinders (HSCs) in isotropic and nematic states using the $NP_nAT$ ensemble Monte Carlo (MC) method in which a normal pressure tensor component is fixed in a system confined between two hard walls. The method allows one to estimate the location of the isotropic-nematic phase transition and to observe the asymmetry in the composition between the coexisting phases, with the expected increase of the HSC concentration in the nematic phase. This is in stark contrast with the previously reported MC simulations where a conventional isotropic $NPT$ ensemble was used. We further compare the simulation results with the theoretical predictions of two analytic theories that extend the original Parsons-Lee theory using the one-fluid and the many-fluid approximation [Malijevský {\it at al} J. Chem. Phys. \textbf{129}, 144504 (2008)]. In the one-fluid version of the theory the properties of the mixture are mapped on an effective one-component HS system while in the many-fluid theory the components of the mixtures are represented as separate effective HS particles. The comparison reveals that both the one- and the many-fluid approaches provide a reasonably accurate quantitative description of the mixture including the predictions of the isotropic-nematic phase boundary and degree of orientational order of the HSC-HS mixtures.
△ Less
Submitted 16 August, 2015;
originally announced August 2015.
-
Directed self-assembly of spherical caps via confinement
Authors:
Carlos Avendano,
Chekesha M. Liddell Watson,
Fernando A. Escobedo
Abstract:
In this work we use Monte Carlo simulations to study the phase behavior of spherical caps confined between two parallel hard walls separated by a distance H. The particle model consists of a hard sphere of diameter σcut off by a plane at a height χ, and it is loosely based on mushroom cap-shaped particles whose phase behavior was recently studied experimentally [E. K. Riley and C. M. Liddell, Lang…
▽ More
In this work we use Monte Carlo simulations to study the phase behavior of spherical caps confined between two parallel hard walls separated by a distance H. The particle model consists of a hard sphere of diameter σcut off by a plane at a height χ, and it is loosely based on mushroom cap-shaped particles whose phase behavior was recently studied experimentally [E. K. Riley and C. M. Liddell, Langmuir, 26, 11648 (2010)]. The geometry of the particles is characterized by the reduced height χ^* = χ/σ, such that the model extrapolates between hard spheres for χ^* \leftarrow 1 and infinitely thin hard platelets for χ^* \letfarrow 0. Three different particle shapes are investigated: (a) three-quarter height spherical caps (χ^* = 3/4), (b) one-half height spherical caps or hemispheres (χ^* = 1/2), and (c) one-quarter height spherical caps (χ^* = 1/4). These three models are used to rationalize the effect of particle shape, obtained by cutting off spheres at different heights, on the entropy-driven self-assembly of the particles under strong confinements; i.e., for 1 < H/χ< 2.5. As H is varied, a sequence of crystal structures are observed, including some having similar symmetry as that of the structures observed in confined hard spheres on account of the remaining spherical surface in the particles, but with additional features on account of the particle shapes having intrinsic anisotropy and orientational degrees of freedom. The χ^* = 3/4 system is found to exhibit a phase diagram that is most similar to the one obtained experimentally for the confined mushroom cap-shaped colloidal particles under. A qualitative global phase diagram is constructed that helps reveal the interrelations among different phases for all the particle shapes and confinements studied.
△ Less
Submitted 2 April, 2013;
originally announced April 2013.
