Source author record

Nicola Ferri

Nicola Ferri appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

eess.AS Artificial Intelligence cond-mat.mes-hall cond-mat.mtrl-sci

Catalog footprint

What is connected

4works

4topics

4close collaborators

Actions

Connect this record

Open graph Browse works

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2022arXiv

Conformer with dual-mode chunked attention for joint online and offline ASR

In this paper, we present an in-depth study on online attention mechanisms and distillation techniques for dual-mode (i.e., joint online and offline) ASR using the Conformer Transducer. In the dual-mode Conformer Transducer model, layers can function in online or offline mode while sharing parameters, and in-place knowledge distillation from offline to online mode is applied in training to improve online accuracy. In our study, we first demonstrate accuracy improvements from using chunked attention in the Conformer encoder compared to autoregressive attention with and without lookahead. Furthermore, we explore the efficient KLD and 1-best KLD losses with different shifts between online and offline outputs in the knowledge distillation. Finally, we show that a simplified dual-mode Conformer that only has mode-specific self-attention performs equally well as the one also having mode-specific convolutions and normalization. Our experiments are based on two very different datasets: the Librispeech task and an internal corpus of medical conversations. Results show that the proposed dual-mode system using chunked attention yields 5% and 4% relative WER improvement on the Librispeech and medical tasks, compared to the dual-mode system using autoregressive attention with similar average lookahead.

preprint2022arXiv

On the Prediction Network Architecture in RNN-T for ASR

RNN-T models have gained popularity in the literature and in commercial systems because of their competitiveness and capability of operating in online streaming mode. In this work, we conduct an extensive study comparing several prediction network architectures for both monotonic and original RNN-T models. We compare 4 types of prediction networks based on a common state-of-the-art Conformer encoder and report results obtained on Librispeech and an internal medical conversation data set. Our study covers both offline batch-mode and online streaming scenarios. In contrast to some previous works, our results show that Transformer does not always outperform LSTM when used as prediction network along with Conformer encoder. Inspired by our scoreboard, we propose a new simple prediction network architecture, N-Concat, that outperforms the others in our on-line streaming benchmark. Transformer and n-gram reduced architectures perform very similarly yet with some important distinct behaviour in terms of previous context. Overall we obtained up to 4.1 % relative WER improvement compared to our LSTM baseline, while reducing prediction network parameters by nearly an order of magnitude (8.4 times).

preprint2016arXiv

Adsorption structures and energetics of molecules on metal surfaces: Bridging experiment and theory

Adsorption geometry and stability of organic molecules on surfaces are key parameters that determine the observable properties and functions of hybrid inorganic/organic systems (HIOSs). Despite many recent advances in precise experimental characterization and improvements in first-principles electronic structure methods, reliable databases of structures and energetics for large adsorbed molecules are largely amiss. In this review, we present such a database for a range of molecules adsorbed on metal single-crystal surfaces. The systems we analyze include noble-gas atoms, conjugated aromatic molecules, carbon nanostructures, and heteroaromatic compounds adsorbed on five different metal surfaces. The overall objective is to establish a diverse benchmark dataset that enables an assessment of current and future electronic structure methods, and motivates further experimental studies that provide ever more reliable data. Specifically, the benchmark structures and energetics from experiment are here compared with the recently developed van der Waals (vdW) inclusive density-functional theory (DFT) method, DFT+vdW$^{\mathrm{surf}}$. In comparison to 23 adsorption heights and 17 adsorption energies from experiment we find a mean average deviation of 0.06 Å and 0.16 eV, respectively. This confirms the DFT+vdW$^{\mathrm{surf}}$ method as an accurate and efficient approach to treat HIOSs. A detailed discussion identifies remaining challenges to be addressed in future development of electronic structure methods, for which the here presented benchmark database may serve as an important reference.

preprint2011arXiv

Carbon sp chains in graphene nanoholes

Nowadays sp carbon chains terminated by graphene or graphitic-like carbon are synthesized routinely in several nanotech labs. We propose an ab-initio study of such carbon-only materials, by computing their structure and stability, as well as their electronic, vibrational and magnetic properties. We adopt a fair compromise of microscopic realism with a certain level of idealization in the model configurations, and predict a number of properties susceptible to comparison with experiment.