Search

Speaker Diarization (SD) aims at grouping speech segments that belong to the same speaker. This task is required in many speech-processing applications, such as rich meeting transcription. In this context, distant microphone arrays usually capture the audio signal. Beamforming, i.e., spatial filtering, is a common practice to process multi-microphone audio data. However, it often requires an explicit localization of the active source to steer the filter. This paper proposes a self-attention-based algorithm to select the output of a bank of fixed spatial filters. This method serves as a feature extractor for joint Voice Activity (VAD) and Overlapped Speech Detection (OSD). The speaker diarization is then inferred from the detected segments. The approach shows convincing distant VAD, OSD, and SD performance, e.g. 14.5% DER on the AISHELL-4 dataset. The analysis of the self-attention weights demonstrates their explainability, as they correlate with the speaker's angular locations., Comment: 5 pages, 2 figures, 2 tables, accepted at Interspeech 2024

The EXO-POPP project aims to establish a comprehensive database comprising 300,000 marriage records from Paris and its suburbs, spanning the years 1880 to 1940, which are preserved in over 130,000 scans of double pages. Each marriage record may encompass up to 118 distinct types of information that require extraction from plain text. In this paper, we introduce the M-POPP dataset, a subset of the M-POPP database with annotations for full-page text recognition and information extraction in both handwritten and printed documents, and which is now publicly available. We present a fully end-to-end architecture adapted from the DAN, designed to perform both handwritten text recognition and information extraction directly from page images without the need for explicit segmentation. We showcase the information extraction capabilities of this architecture by achieving a new state of the art for full-page Information Extraction on Esposalles and we use this architecture as a baseline for the M-POPP dataset. We also assess and compare how different encoding strategies for named entities in the text affect the performance of jointly recognizing handwritten text and extracting information, from full pages., Comment: To be published in: International Conference on Document Analysis and Recognition - ICDAR 2024

This paper presents a semi-automatic approach to create a diachronic corpus of voices balanced for speaker's age, gender, and recording period, according to 32 categories (2 genders, 4 age ranges and 4 recording periods). Corpora were selected at French National Institute of Audiovisual (INA) to obtain at least 30 speakers per category (a total of 960 speakers; only 874 have be found yet). For each speaker, speech excerpts were extracted from audiovisual documents using an automatic pipeline consisting of speech detection, background music and overlapped speech removal and speaker diarization, used to present clean speaker segments to human annotators identifying target speakers. This pipeline proved highly effective, cutting down manual processing by a factor of ten. Evaluation of the quality of the automatic processing and of the final output is provided. It shows the automatic processing compare to up-to-date process, and that the output provides high quality speech for most of the selected excerpts. This method shows promise for creating large corpora of known target speakers., Comment: Keywords:, semi-automatic processing, corpus creation, diarization, speaker identification, gender-balanced, age-balanced, speaker corpus, diachrony

Voice Activity Detection (VAD) and Overlapped Speech Detection (OSD) are key pre-processing tasks for speaker diarization. In the meeting context, it is often easier to capture speech with a distant device. This consideration however leads to severe performance degradation. We study a unified supervised learning framework to solve distant multi-microphone joint VAD and OSD (VAD+OSD). This paper investigates various multi-channel VAD+OSD front-ends that weight and combine incoming channels. We propose three algorithms based on the Self-Attention Channel Combinator (SACC), previously proposed in the literature. Experiments conducted on the AMI meeting corpus exhibit that channel combination approaches bring significant VAD+OSD improvements in the distant speech scenario. Specifically, we explore the use of learned complex combination weights and demonstrate the benefits of such an approach in terms of explainability. Channel combination-based VAD+OSD systems are evaluated on the final back-end task, i.e. speaker diarization, and show significant improvements. Finally, since multi-channel systems are trained given a fixed array configuration, they may fail in generalizing to other array set-ups, e.g. mismatched number of microphones. A channel-number invariant loss is proposed to learn a unique feature representation regardless of the number of available microphones. The evaluation conducted on mismatched array configurations highlights the robustness of this training strategy., Comment: 14 pages, 5 figures, accepted at IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP)

The study of hardest and easiest fitness landscapes is an active area of research. Recently, Kaufmann, Larcher, Lengler and Zou conjectured that for the self-adjusting $(1,\lambda)$-EA, Adversarial Dynamic BinVal (ADBV) is the hardest dynamic monotone function to optimize. We introduce the function Switching Dynamic BinVal (SDBV) which coincides with ADBV whenever the number of remaining zeros in the search point is strictly less than $n/2$, where $n$ denotes the dimension of the search space. We show, using a combinatorial argument, that for the $(1+1)$-EA with any mutation rate $p \in [0,1]$, SDBV is drift-minimizing among the class of dynamic monotone functions. Our construction provides the first explicit example of an instance of the partially-ordered evolutionary algorithm (PO-EA) model with parameterized pessimism introduced by Colin, Doerr and F\'erey, building on work of Jansen. We further show that the $(1+1)$-EA optimizes SDBV in $\Theta(n^{3/2})$ generations. Our simulations demonstrate matching runtimes for both static and self-adjusting $(1,\lambda)$ and $(1+\lambda)$-EA. We further show, using an example of fixed dimension, that drift-minimization does not equal maximal runtime.

Recent architectural developments have enabled recurrent neural networks (RNNs) to reach and even surpass the performance of Transformers on certain sequence modeling tasks. These modern RNNs feature a prominent design pattern: linear recurrent layers interconnected by feedforward paths with multiplicative gating. Here, we show how RNNs equipped with these two design elements can exactly implement (linear) self-attention, the main building block of Transformers. By reverse-engineering a set of trained RNNs, we find that gradient descent in practice discovers our construction. In particular, we examine RNNs trained to solve simple in-context learning tasks on which Transformers are known to excel and find that gradient descent instills in our RNNs the same attention-based in-context learning algorithm used by Transformers. Our findings highlight the importance of multiplicative interactions in neural networks and suggest that certain RNNs might be unexpectedly implementing attention under the hood.

The strategy of training the model from scratch in a specific language or domain serves two essential purposes: i) enhancing performance in the particular linguistic or domain context, and ii) ensuring effective tokenization. The main limitation inherent to this approach lies in the associated cost, which can reach six to seven-digit dollar values, depending on the model size and the number of parameters involved. The main solution to overcome the cost challenge is to rely on available pre-trained models, which, despite recent advancements such as the LLaMA and LLaMA-2 models, still demonstrate inefficiency for certain specific domain problems or prove ineffective in scenarios involving conversational memory resources, given the large number of tokens required to represent text. To overcome this issue, we present a methodology named Cabrita, which, as our research demonstrates, successfully addresses the performance and efficient tokenization problem, all at an affordable cost. We believe that this methodology can be applied to any transformer-like architecture model. To validate the study, we conducted continuous pre-training exclusively using Portuguese text on a 3-billion-parameter model known as OpenLLaMA, resulting in a model named openCabrita 3B. The openCabrita 3B also features a new tokenizer that results in a significant reduction in the number of tokens required to represent the text. In our assessment, for few-shot learning tasks, we achieved similar results with this 3B model compared to a traditional continuous pre-training approach as well as to 7B models English pre-trained models., Comment: 9 pages, 1 figure

Computational fluid dynamics is intensively used to deepen the understanding of aneurysm growth and rupture in the attempt to support physicians during therapy planning. Numerous studies have assumed fully-rigid vessel walls in their simulations, whose sole hemodynamics may fail to provide a satisfactory criterion for rupture risk assessment. Moreover, direct in-vivo observations of intracranial aneurysm pulsation have been recently reported, encouraging the development of fluid-structure interaction for their modelling and for new assessments. In this work, we describe a new fluid-structure interaction benchmark setting for the careful evaluation of different aneurysm shapes. The studied configurations consist of three real aneurysm domes positioned on a toroidal channel. All geometric features, meshing characteristics, flow quantities, comparisons with a rigid-wall model and corresponding plots are provided. Reported results emphasize the alteration of flow patterns and hemodynamic descriptors when moving from the rigid-wall model to the complete fluid-structure interaction framework, thereby underlining the importance of the coupling between hemodynamics and the surrounding vessel tissue., Comment: 23 pages, 14 figures

The difficulty to measure or predict species community composition at fine spatio-temporal resolution and over large spatial scales severely hampers our ability to understand species assemblages and take appropriate conservation measures. Despite the progress in species distribution modeling (SDM) over the past decades, SDM have just begun to integrate high resolution remote sensing data and their predictions are still entailed by many biases due to heterogeneity of the available biodiversity observations, most often opportunistic presence only data. We designed a European scale dataset covering around ten thousand plant species to calibrate and evaluate SDM predictions of species composition in space and time at high spatial resolution (~ten meters), and their spatial transferability. For model training, we extracted and harmonized five million heterogeneous presence-only records from selected GBIF datasets and 6 thousand exhaustive presence-absence surveys both sampled during 2017-2021. We associated species observations to diverse environmental rasters classically used in SDMs, as well as to 10 m resolution RGB and Near-Infra-Red satellite images and 20 years-time series of climatic variables and satellite point values. The evaluation dataset is based on 22 thousand standardized presence-absence surveys separated from the training set with a spatial block hold out procedure. The GeoLifeCLEF 2023 dataset is open access and the first benchmark for researchers aiming to improve the prediction of plant species composition at a very fine spatial grain and at continental scale. It is a space to explore new ways of combining massive and diverse species observations and environmental information at various scales. Innovative AI-based approaches, in particular, should be among the most interesting methods to experiment with on the GeoLifeCLEF 2023 dataset., Comment: The full dataset is freely available at the link below (perennial repository) for academic use or other non-commercial use: https://lab.plantnet.org/seafile/d/936fe4298a5a4f4c8dbd/

Voice activity and overlapped speech detection (respectively VAD and OSD) are key pre-processing tasks for speaker diarization. The final segmentation performance highly relies on the robustness of these sub-tasks. Recent studies have shown VAD and OSD can be trained jointly using a multi-class classification model. However, these works are often restricted to a specific speech domain, lacking information about the generalization capacities of the systems. This paper proposes a complete and new benchmark of different VAD and OSD models, on multiple audio setups (single/multi-channel) and speech domains (e.g. media, meeting...). Our 2/3-class systems, which combine a Temporal Convolutional Network with speech representations adapted to the setup, outperform state-of-the-art results. We show that the joint training of these two tasks offers similar performances in terms of F1-score to two dedicated VAD and OSD systems while reducing the training cost. This unique architecture can also be used for single and multichannel speech processing.

Speaker diarization is the task of answering Who spoke and when? in an audio stream. Pipeline systems rely on speech segmentation to extract speakers' segments and achieve robust speaker diarization. This paper proposes a common framework to solve three segmentation tasks in the distant speech scenario: Voice Activity Detection (VAD), Overlapped Speech Detection (OSD), and Speaker Change Detection (SCD). In the literature, a few studies investigate the multi-microphone distant speech scenario. In this work, we propose a new set of spatial features based on direction-of-arrival estimations in the circular harmonic domain (CH-DOA). These spatial features are extracted from multi-microphone audio data and combined with standard acoustic features. Experiments on the AMI meeting corpus show that CH-DOA can improve the segmentation while being robust in the case of deactivated microphones., Comment: Interspeech 2023, international Speech Communication Association (ISCA), Aug 2023, Dublin, Ireland

Text summarization is a downstream natural language processing (NLP) task that challenges the understanding and generation capabilities of language models. Considerable progress has been made in automatically summarizing short texts, such as news articles, often leading to satisfactory results. However, summarizing long documents remains a major challenge. This is due to the complex contextual information in the text and the lack of open-source benchmarking datasets and evaluation frameworks that can be used to develop and test model performance. In this work, we use ChatGPT, the latest breakthrough in the field of large language models (LLMs), together with the extractive summarization model C2F-FAR (Coarse-to-Fine Facet-Aware Ranking) to propose a hybrid extraction and summarization pipeline for long documents such as business articles and books. We work with the world-renowned company getAbstract AG and leverage their expertise and experience in professional book summarization. A practical study has shown that machine-generated summaries can perform at least as well as human-written summaries when evaluated using current automated evaluation metrics. However, a closer examination of the texts generated by ChatGPT through human evaluations has shown that there are still critical issues in terms of text coherence, faithfulness, and style. Overall, our results show that the use of ChatGPT is a very promising but not yet mature approach for summarizing long documents and can at best serve as an inspiration for human editors. We anticipate that our work will inform NLP researchers about the extent to which ChatGPT's capabilities for summarizing long documents overlap with practitioners' needs. Further work is needed to test the proposed hybrid summarization pipeline, in particular involving GPT-4, and to propose a new evaluation framework tailored to the task of summarizing long documents.

Speech data carries a range of personal information, such as the speaker's identity and emotional state. These attributes can be used for malicious purposes. With the development of virtual assistants, a new generation of privacy threats has emerged. Current studies have addressed the topic of preserving speech privacy. One of them, the VoicePrivacy initiative aims to promote the development of privacy preservation tools for speech technology. The task selected for the VoicePrivacy 2020 Challenge (VPC) is about speaker anonymization. The goal is to hide the source speaker's identity while preserving the linguistic information. The baseline of the VPC makes use of a voice conversion. This paper studies the impact of the speaker anonymization baseline system of the VPC on emotional information present in speech utterances. Evaluation is performed following the VPC rules regarding the attackers' knowledge about the anonymization system. Our results show that the VPC baseline system does not suppress speakers' emotions against informed attackers. When comparing anonymized speech to original speech, the emotion recognition performance is degraded by 15\% relative to IEMOCAP data, similar to the degradation observed for automatic speech recognition used to evaluate the preservation of the linguistic information.

In recent years, large-scale numerical simulations played an essential role in estimating the effects of explosion events in urban environments, for the purpose of ensuring the security and safety of cities. Such simulations are computationally expensive and, often, the time taken for one single computation is large and does not permit parametric studies. The aim of this work is therefore to facilitate real-time and multi-query calculations by employing a non-intrusive Reduced Order Method (ROM). We propose a deep learning-based (DL) ROM scheme able to deal with fast transient dynamics. In the case of blast waves, the parametrised PDEs are time-dependent and non-linear. For such problems, the Proper Orthogonal Decomposition (POD), which relies on a linear superposition of modes, cannot approximate the solutions efficiently. The piecewise POD-DL scheme developed here is a local ROM based on time-domain partitioning and a first dimensionality reduction obtained through the POD. Autoencoders are used as a second and non-linear dimensionality reduction. The latent space obtained is then reconstructed from the time and parameter space through deep forward neural networks. The proposed scheme is applied to an example consisting of a blast wave propagating in air and impacting on the outside of a building. The efficiency of the deep learning-based ROM in approximating the time-dependent pressure field is shown.

Our paper proposes an innovative approach for modeling Fluid-Structure Interaction (FSI). Our method combines both traditional monolithic and partitioned approaches, creating a hybrid solution that facilitates FSI. At each time iteration, the solid mesh is immersed within a fluid-solid mesh, all while maintaining its independent Lagrangian hyperelastic solver. The Eulerian mesh encompasses both the fluid and solid components and accommodates various physical phenomena. We enhance the interaction between solid and fluid through anisotropic mesh adaptation and the Level-Set methods. This enables a more accurate representation of their interaction. Together, these components constitute the Adaptive Immersed Mesh Method (AIMM). For both solvers, we utilize the Variational Multi-Scale (VMS) method, mitigating potential spurious oscillations common with piecewise linear tetrahedral elements. The framework operates in 3D with parallel computing capabilities. Our methods accuracy, robustness, and capabilities are assessed through a series of 2D numerical problems. Furthermore, we present various three-dimensional test cases and compare their results to experimental data.

Mordechay B. Levin has constructed a number $\lambda$ which is normal in base 2, and such that the sequence $(\left\{2^n \lambda\right\})_{n=0,1,2,\ldots}$ has very small discrepancy $D_N$. Indeed we have $N\cdot D_N = \mathcal{O} \left(\left(\log N\right)^2\right)$. This construction technique of Levin was generalized by Becher and Carton, who generated normal numbers via perfect nested necklaces, and they showed that for these normal numbers the same upper discrepancy estimate holds as for the special example of Levin. In this paper now we derive an upper discrepancy bound for so-called semi-perfect nested necklaces and show that for the Levin's normal number in arbitrary prime base $p$ this upper bound for the discrepancy is best possible, i.e., $N\cdot D_N \geq c\left(\log N\right)^2$ with $c>0$ for infinitely many $N$. This result generalizes a previous result where we ensured for the special example of Levin for the base $p=2$, that $N\cdot D_N =O( \left(\log N\right)^2)$ is best possible in $N$. So far it is known by a celebrated result of Schmidt that for any sequence in $[0,1)$, $N\cdot D_N\geq c \log N$ with $c>0$ for infinitely many $N$. So there is a gap of a $\log N$ factor in the question, what is the best order for the discrepancy in $N$ that can be achieved for a normal number. Our result for Levin's normal number in any prime base on the one hand might support the guess that $O( \left(\log N\right)^2)$ is the best order in $N$ that can be achieved by a normal number, while generalizing the class of known normal numbers by introducing e.g. semi-perfect necklaces on the other hand might help for the search of normal numbers that satisfy smaller discrepancy bounds in $N$ than $N\cdot D_N=O( \left(\log N\right)^2)$., Comment: 29 pages

In this working paper we present our current progress in the training of machine learning models to execute short option strategies on the S&P500. As a first step, this paper is breaking this problem down to a supervised classification task to decide if a short straddle on the S&P500 should be executed or not on a daily basis. We describe our used framework and present an overview over our evaluation metrics on different classification models. In this preliminary work, using standard machine learning techniques and without hyperparameter search, we find no statistically significant outperformance to a simple "trade always" strategy, but gain additional insights on how we could proceed in further experiments., Comment: 25 pages

In this letter, we discuss how flow inhomogeneity affects the self-diffusion behavior in granular flows. Whereas self-diffusion scalings have been well characterized in the past for homogeneous shearing, the effect of shear localization and nonlocality of the flow has not been studied. We therefore present measurements of self-diffusion coefficients in discrete numerical simulations of steady, inhomogeneous, and collisional shearing flows of nearly identical, frictional, and inelastic spheres. We focus on a wide range of dense solid volume fractions, that correspond to geophysical and industrial shearing flows that are dominated by collisional interactions. We compare the measured values first, with a scaling based on shear rate and, then, on a scaling based on the granular temperature. We find that the latter does much better than the former in collapsing the data. The results lay the foundations of diffusion models for inhomogeneous shearing flows, which should be useful in treating problems of mixing and segregation.

Speech signals contain a lot of sensitive information, such as the speaker's identity, which raises privacy concerns when speech data get collected. Speaker anonymization aims to transform a speech signal to remove the source speaker's identity while leaving the spoken content unchanged. Current methods perform the transformation by relying on content/speaker disentanglement and voice conversion. Usually, an acoustic model from an automatic speech recognition system extracts the content representation while an x-vector system extracts the speaker representation. Prior work has shown that the extracted features are not perfectly disentangled. This paper tackles how to improve features disentanglement, and thus the converted anonymized speech. We propose enhancing the disentanglement by removing speaker information from the acoustic model using vector quantization. Evaluation done using the VoicePrivacy 2022 toolkit showed that vector quantization helps conceal the original speaker identity while maintaining utility for speech recognition.

Mordechay Levin has constructed a number $\alpha$ which is normal in base 2, and such that the sequence $\left\{2^n \alpha\right\}_{n=0,1,2,\ldots}$ has very small discrepancy $D_N$. Indeed we have $N\cdot D_N = \mathcal{O} \left(\left(\log N\right)^2\right)$. That means, that $\alpha$ is normal of extremely high quality. In this paper we show that this estimate is best possible, i.e., $N\cdot D_N \geq c \cdot \left(\log N\right)^2$ for infinitely many $N$., Comment: There was a gap in the proof in a previous version. Here now is the corrected version

We study the $(1:s+1)$ success rule for controlling the population size of the $(1,\lambda)$-EA. It was shown by Hevia Fajardo and Sudholt that this parameter control mechanism can run into problems for large $s$ if the fitness landscape is too easy. They conjectured that this problem is worst for the OneMax benchmark, since in some well-established sense OneMax is known to be the easiest fitness landscape. In this paper we disprove this conjecture and show that OneMax is not the easiest fitness landscape with respect to finding improving steps. As a consequence, we show that there exists $s$ and $\varepsilon$ such that the self-adjusting $(1,\lambda)$-EA with $(1:s+1)$-rule optimizes OneMax efficiently when started with $\varepsilon n$ zero-bits, but does not find the optimum in polynomial time on Dynamic BinVal. Hence, we show that there are landscapes where the problem of the $(1:s+1)$-rule for controlling the population size of the $(1, \lambda)$-EA is more severe than for OneMax.

We study the $(1,\lambda)$-EA with mutation rate $c/n$ for $c\le 1$, where the population size is adaptively controlled with the $(1:s+1)$-success rule. Recently, Hevia Fajardo and Sudholt have shown that this setup with $c=1$ is efficient on \onemax for $s<1$, but inefficient if $s \ge 18$. Surprisingly, the hardest part is not close to the optimum, but rather at linear distance. We show that this behavior is not specific to \onemax. If $s$ is small, then the algorithm is efficient on all monotone functions, and if $s$ is large, then it needs superpolynomial time on all monotone functions. In the former case, for $c<1$ we show a $O(n)$ upper bound for the number of generations and $O(n\log n)$ for the number of function evaluations, and for $c=1$ we show $O(n\log n)$ generations and $O(n^2\log\log n)$ evaluations. We also show formally that optimization is always fast, regardless of $s$, if the algorithm starts in proximity of the optimum. All results also hold in a dynamic environment where the fitness function changes in each generation.