arXiv subjects

Marc Delcroix

Publications and source records attributed to Marc Delcroix.

119 records · Page 7Linked to original sources

arXiv2020

Tackling real noisy reverberant meetings with all-neural source separation, counting, and diarization system

Automatic meeting analysis is an essential fundamental technology required to let, e.g. smart devices follow and respond to our conversations. To achieve an optimal automatic meeting analysis, we previously proposed an all-neural approach that jointly solves source separation, speaker diarization and source counting problems in an optimal way (in a sense that all the 3 tasks can be jointly optimized through error back-propagation). It was shown that the method could well handle simulated clean (noiseless and anechoic) dialog-like data, and achieved very good performance in comparison with several conventional methods. However, it was not clear whether such all-neural approach would be successfully generalized to more complicated real meeting data containing more spontaneously-speaking speakers, severe noise and reverberation, and how it performs in comparison with the state-of-the-art systems in such scenarios. In this paper, we first consider practical issues required for improving the robustness of the all-neural approach, and then experimentally show that, even in real meeting scenarios, the all-neural approach can perform effective speech enhancement, and simultaneously outperform state-of-the-art systems.

eess.AS↗

arXiv2020

Improving noise robust automatic speech recognition with single-channel time-domain enhancement network

With the advent of deep learning, research on noise-robust automatic speech recognition (ASR) has progressed rapidly. However, ASR performance in noisy conditions of single-channel systems remains unsatisfactory. Indeed, most single-channel speech enhancement (SE) methods (denoising) have brought only limited performance gains over state-of-the-art ASR back-end trained on multi-condition training data. Recently, there has been much research on neural network-based SE methods working in the time-domain showing levels of performance never attained before. However, it has not been established whether the high enhancement performance achieved by such time-domain approaches could be translated into ASR. In this paper, we show that a single-channel time-domain denoising approach can significantly improve ASR performance, providing more than 30 % relative word error reduction over a strong ASR back-end on the real evaluation data of the single-channel track of the CHiME-4 dataset. These positive results demonstrate that single-channel noise reduction can still improve ASR performance, which should open the door to more research in that direction.

eess.AS↗

arXiv2020

Fragmentation modelling of the August 2019 impact on Jupiter

On 7th August 2019, an impact flash lasting $\sim1$s was observed on Jupiter. The video of this event was analysed to obtain the lightcurve and determine the energy release and initial mass. We find that the impactor released a total energy of $96-151$ kilotons of TNT, corresponding to an initial mass between $190-260$ metric tonnes with a diameter between $4-10$m. We developed a fragmentation model to simulate the atmospheric breakup of the object and reproduce the lightcurve. We model three different materials: cometary, stony and metallic at speeds of $60$, $65 $ and $70$ km/s to determine the material makeup of the impacting object. The slower cases are best fit by a strong, metallic object while the faster cases require a weaker material.

astro-ph.EP↗

arXiv2020

Speech Enhancement using Self-Adaptation and Multi-Head Self-Attention

This paper investigates a self-adaptation method for speech enhancement using auxiliary speaker-aware features; we extract a speaker representation used for adaptation directly from the test utterance. Conventional studies of deep neural network (DNN)--based speech enhancement mainly focus on building a speaker independent model. Meanwhile, in speech applications including speech recognition and synthesis, it is known that model adaptation to the target speaker improves the accuracy. Our research question is whether a DNN for speech enhancement can be adopted to unknown speakers without any auxiliary guidance signal in test-phase. To achieve this, we adopt multi-task learning of speech enhancement and speaker identification, and use the output of the final hidden layer of speaker identification branch as an auxiliary feature. In addition, we use multi-head self-attention for capturing long-term dependencies in the speech and noise. Experimental results on a public dataset show that our strategy achieves the state-of-the-art performance and also outperform conventional methods in terms of subjective quality.

eess.AS↗

arXiv2020

Improving speaker discrimination of target speech extraction with time-domain SpeakerBeam

Target speech extraction, which extracts a single target source in a mixture given clues about the target speaker, has attracted increasing attention. We have recently proposed SpeakerBeam, which exploits an adaptation utterance of the target speaker to extract his/her voice characteristics that are then used to guide a neural network towards extracting speech of that speaker. SpeakerBeam presents a practical alternative to speech separation as it enables tracking speech of a target speaker across utterances, and achieves promising speech extraction performance. However, it sometimes fails when speakers have similar voice characteristics, such as in same-gender mixtures, because it is difficult to discriminate the target speaker from the interfering speakers. In this paper, we investigate strategies for improving the speaker discrimination capability of SpeakerBeam. First, we propose a time-domain implementation of SpeakerBeam similar to that proposed for a time-domain audio separation network (TasNet), which has achieved state-of-the-art performance for speech separation. Besides, we investigate (1) the use of spatial features to better discriminate speakers when microphone array recordings are available, (2) adding an auxiliary speaker identification loss for helping to learn more discriminative voice characteristics. We show experimentally that these strategies greatly improve speech extraction performance, especially for same-gender mixtures, and outperform TasNet in terms of target speech extraction.

eess.AS↗

arXiv2019

The onset and growth of the 2018 Martian Global Dust Storm

We analyze the onset and initial expansion of the 2018 Martian Global Dust Storm (GDS 2018) using ground-based images in the visual range. This is the first case of a confirmed GDS initiating in the Northern Hemisphere. A dusty area extending about 1.4x10e5 km^2 and centered at latitude +31.7° $\pm$ 1.8° and west longitude 18° $\pm$ 5°W in Acidalia Planitia was captured on 30 and 31 May 2018 (Ls = 184.9°). From 1 to 8 June, daily image series showed the storm expanding southwards along the Acidalia corridor with velocities of 5 m/s, and simultaneously progressing eastwards and westwards with horizontal velocities ranging from 5 to 40 m/s. By 8 June the dust reached latitude -55° and later penetrated in the South polar region, whereas in the North the dust progression stopped at latitude +46°. We compare the onset and expansion stage of this GDS with the previous confirmed storms.

astro-ph.EP↗

arXiv2019

All-neural online source separation, counting, and diarization for meeting analysis

Automatic meeting analysis comprises the tasks of speaker counting, speaker diarization, and the separation of overlapped speech, followed by automatic speech recognition. This all has to be carried out on arbitrarily long sessions and, ideally, in an online or block-online manner. While significant progress has been made on individual tasks, this paper presents for the first time an all-neural approach to simultaneous speaker counting, diarization and source separation. The NN-based estimator operates in a block-online fashion and tracks speakers even if they remain silent for a number of time blocks, thus learning a stable output order for the separated sources. The neural network is recurrent over time as well as over the number of sources. The simulation experiments show that state of the art separation performance is achieved, while at the same time delivering good diarization and source counting results. It even generalizes well to an unseen large number of blocks.

eess.AS↗

arXiv2018

Analysis of Neptune's 2017 Bright Equatorial Storm

We report the discovery of a large ($\sim$8500 km diameter) infrared-bright storm at Neptune's equator in June 2017. We tracked the storm over a period of 7 months with high-cadence infrared snapshot imaging, carried out on 14 nights at the 10 meter Keck II telescope and 17 nights at the Shane 120 inch reflector at Lick Observatory. The cloud feature was larger and more persistent than any equatorial clouds seen before on Neptune, remaining intermittently active from at least 10 June to 31 December 2017. Our Keck and Lick observations were augmented by very high-cadence images from the amateur community, which permitted the determination of accurate drift rates for the cloud feature. Its zonal drift speed was variable from 10 June to at least 25 July, but remained a constant $237.4 \pm 0.2$ m s$^{-1}$ from 30 September until at least 15 November. The pressure of the cloud top was determined from radiative transfer calculations to be 0.3-0.6 bar; this value remained constant over the course of the observations. Multiple cloud break-up events, in which a bright cloud band wrapped around Neptune's equator, were observed over the course of our observations. No "dark spot" vortices were seen near the equator in HST imaging on 6 and 7 October. The size and pressure of the storm are consistent with moist convection or a planetary-scale wave as the energy source of convective upwelling, but more modeling is required to determine the driver of this equatorial disturbance as well as the triggers for and dynamics of the observed cloud break-up events.

astro-ph.EP↗

arXiv2018

A New, Long-Lived, Jupiter Mesoscale Wave Observed at Visible Wavelengths

Small-scale waves were observed along the boundary between Jupiter's North Equatorial Belt and North Tropical Zone, ~16.5° N planetographic latitude in Hubble Space Telescope data in 2012 and throughout 2015 to 2018, observable at all wavelengths from the UV to the near IR. At peak visibility, the waves have sufficient contrast (~10%) to be observed from ground-based telescopes. They have a typical wavelength of about 1.2° (1400 km), variable-length wave trains, and westward phase speeds of a few m/s or less. New analysis of Voyager 2 data shows similar wave trains over at least 300 hours. Some waves appear curved when over cyclones and anticyclones, but most are straight, but tilted, shifting in latitude as they pass vortices. Based on their wavelengths, phase speeds, and faint appearance at high-altitude sensitive passbands, the observed NEB waves are consistent with inertia-gravity waves at the 500-mbar pressure level, though formation altitude is not well constrained. Preliminary General Circulation Model simulations generate inertia-gravity waves from vortices interacting with the environment and can reproduce the observed wavelengths and orientations. Several mechanisms can generate these waves, and all may contribute: geostrophic adjustment of cyclones; cyclone/anticyclone interactions; wind interactions with obstructions or heat pulses from convection; or changing vertical wind shear. However, observations also show that the presence of vortices and/or regions of convection are not sufficient by themselves for wave formation, implying that a change in vertical structure may affect their stability, or that changes in haze properties may affect their visibility.

astro-ph.EP↗

arXiv2015

The need for Professional-Amateur collaborations in studies of Jupiter and Saturn

The observation of gaseous giant planets is of high scientific interest. Although they have been the targets of several spacecraft missions, there still remains a need for continuous ground-based observations. As their atmospheres present fast dynamic environments on various time scales, the availability of time at professional telescopes is neither uniform nor of sufficient duration to assess temporal changes. However, numerous amateurs with small telescopes (of 15-40 cm) and modern hardware and software equipment can monitor these changes daily (within the 360-900nm range). Amateurs are able to trace the structure and the evolution of atmospheric features, such as major planetary-scale disturbances, vortices, and storms. Their observations provide a continuous record and it is not uncommon to trigger professional observations in cases of important events, such as sudden onset of global changes, storms and celestial impacts. For example, the continuous amateur monitoring has led to the discovery of fireballs in Jupiter's atmosphere, providing information not only on Jupiter's gravitational influence but also on the properties and populations of the impactors. Photometric monitoring of stellar occultations by the planets can reveal spatial/temporal variability in their atmospheric structure. Therefore, co-ordination and communication between professionals and amateurs is important. We present examples of such collaborations that: (i) engage systematic multi-wavelength observations and databases, (ii) examine the variability of cloud features over timescales from days to decades, (iii) provide, by ground-based professional and amateur observations, the necessary spatial and temporal resolution of features that will be studied by the interplanetary mission Juno, (iv) investigate video observations of Jupiter to identify impacts of small objects, (v) carry out stellar-occultation campaigns.

astro-ph.EP↗

arXiv2010

Contribution of amateur observations to Saturn storm studies

Since 2004, Saturn Electrostatic Discharges (SEDs), which are the radio signatures of lightning in Saturn's atmosphere, have been observed by the Cassini Radio and Plasma Wave Science instrument (RPWS). Despite their important time coverage, these observations lack the resolution and positioning given by imaging around visible wavelengths. Amateur observations from Earth have been increasing in quality and coverage since a few years, bringing information on positions, drift rates and shape evolutions of large visible white spots in Saturn's atmosphere. Combining these two complementary sources has brought better analysis of Saturn's storms evolutions.

astro-ph.EP↗