arXiv subjects

Marcin Witkowski

Publications and source records attributed to Marcin Witkowski.

17 recordsLinked to original sources

arXiv2026

HALAS: A Human-Annotated Dataset of Hallucinations of Modern ASR Systems

End-to-end Automatic Speech Recognition (ASR) systems hallucinate on natural speech, yet existing mitigation methods are typically evaluated on non-speech or artificially corrupted audio. We introduce HALAS, the first human-annotated dataset of naturally occurring hallucinations from seven state-of-the-art ASR models on real unprocessed earnings call recordings. HALAS provides span-level labels, enabling analysis of hallucination patterns and their severity. Our analysis reveals strong cross-model vocabulary overlap and confirms that hallucinations also occur for almost correctly transcribed speech (characterized by a low Word Error Rate). The proposed benchmark with HALAS shows that the character and semantic-level metrics used as a proxy for hallucination detection reach 81% ROC-AUC, while state-of-the-art detection methods achieve an F1 score of only 53.1%. As such, HALAS establishes the first rigorous non-artificial benchmark for the detection and mitigation of ASR hallucinations.

cs.SD↗

arXiv2026

From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection

Hallucinations of ASR models - fluent transcriptions with no basis in audio - degrade system performance and pose risks in downstream applications. Robust detection of such errors remains a challenge. This paper studies Whisper large v3 hallucination detection on real-speech human-annotated data across three paradigms: text-based, LLM-based, and internal decoder state probing. Text classifiers utilizing metrics for text evaluation achieve high recall but degrade without reference transcripts. LLM-based detection improves precision with domain-specific prompt conditioning, yet remains less competitive than the lightweight text-based methods. Probing Whisper's decoder representations, without a ground-truth reference, yields the strongest performance, revealing that hallucination traits are encoded across intermediate decoding layers. A late-fusion meta-classifier combining text and internal-state outputs achieves the best overall detection performance.

cs.SD↗

arXiv2026

Isotopic effect on collisional widths and shifts of Hg clock transition induced by cold Rb atoms

We study the isotopic dependence of collisional widths and shifts of the Hg clock transition $^1$S$_0$-$^3$P$_0$ perturbed by the Rb atoms in the temperature range from 1 nK to 1 K. For this purpose, we model the Born-Oppenheimer effective interaction potential by including the leading long-range van der Waals coefficients. For elastic collisions, we show the connection between the dependence of collision line shape parameters on the reduced mass of colliding partners as well as the variation of the scattering length in the excited and ground states of the Hg-Rb system in the $\upmu$K temperature range. We confront the full quantum scattering calculations with a semi-classical approximation for collisional widths and shifts. We show that the shape resonances in excited and ground scattering states lead to significant variations of collisional line shape parameters with the change of the reduced mass of colliding atoms. We also indicate the possible influence of inelastic collisions, which could lead to universal behavior and significantly affect the dependence of collisional broadening and shifting on the isotopic combination of colliding atoms.

physics.atom-ph↗

arXiv2025

Investigation of Whisper ASR Hallucinations Induced by Non-Speech Audio

Hallucinations of deep neural models are amongst key challenges in automatic speech recognition (ASR). In this paper, we investigate hallucinations of the Whisper ASR model induced by non-speech audio segments present during inference. By inducting hallucinations with various types of sounds, we show that there exists a set of hallucinations that appear frequently. We then study hallucinations caused by the augmentation of speech with such sounds. Finally, we describe the creation of a bag of hallucinations (BoH) that allows to remove the effect of hallucinations through the post-processing of text transcriptions. The results of our experiments show that such post-processing is capable of reducing word error rate (WER) and acts as a good safeguard against problematic hallucinations.

cs.SD↗

arXiv2024

Half-integral Erd\H{o}s-P\'{o}sa property for non-null $S$-$T$ paths

For a group $\Gamma$, a $\Gamma$-labelled graph is an undirected graph $G$ where every orientation of an edge is assigned an element of $\Gamma$ so that opposite orientations of the same edge are assigned inverse elements. A path in $G$ is non-null if the product of the labels along the path is not the neutral element of $\Gamma$. We prove that for every finite group $\Gamma$, non-null $S$-$T$ paths in $\Gamma$-labelled graphs exhibit the half-integral Erd\H{o}s-P\'osa property. More precisely, there is a function $f$, depending on $\Gamma$, such that for every $\Gamma$-labelled graph $G$, subsets of vertices $S$ and $T$, and integer $k$, one of the following objects exists: a family $\cal F$ consisting of $k$ non-null $S$-$T$ paths in $G$ such that every vertex of $G$ participates in at most two paths of $\cal F$; or a set $X$ consisting of at most $f(k)$ vertices that meets every non-null $S$-$T$ path in $G$. This in particular proves that in undirected graphs $S$-$T$ paths of odd length have the half-integral Erd\H{o}s-P\'osa property.

math.CO↗

arXiv2023

BUT CHiME-7 system description

This paper describes the joint effort of Brno University of Technology (BUT), AGH University of Krakow and University of Buenos Aires on the development of Automatic Speech Recognition systems for the CHiME-7 Challenge. We train and evaluate various end-to-end models with several toolkits. We heavily relied on Guided Source Separation (GSS) to convert multi-channel audio to single channel. The ASR is leveraging speech representations from models pre-trained by self-supervised learning, and we do a fusion of several ASR systems. In addition, we modified external data from the LibriSpeech corpus to become a close domain and added it to the training. Our efforts were focused on the far-field acoustic robustness sub-track of Task 1 - Distant Automatic Speech Recognition (DASR), our systems use oracle segmentation.

cs.SD↗

arXiv2022

Absolute frequency measurement of the $6s^2~^1S_0 \to 6s6p~^3P_1$ $F=3/2\to F'=5/2$ $^{\text{201}}$Hg transition with background-free saturation spectroscopy

We report the development of a method for eliminating background-induced systematic shifts affecting precise measurements of saturation absorption signals. With this technique, we measured the absolute frequency of the $6s^2~^1\text{S}_0 \to 6s6p~^3\text{P}_1$ transition in $^{201a}\text{Hg}$ ($F=3/2\to F'=5/2$) to be $1181541111051(83)$~kHz. The measurement was referenced with an optical frequency comb synchronized to the frequency of the local representation of the UTC. This specific atomic line is situated on the steep slope of the Doppler background at room temperature, which results in frequency systematic shift. We determined the dependence of this shift on the properties of both the spectral line and the background of the measured signal.

physics.atom-ph↗

arXiv2022

Photoionization cross sections of ultracold $^{88}$Sr in $^1$P$_1$ and $^3$S$_1$ states at 390 nm and the resulting blue-detuned magic wavelength optical lattice clock constraints

We present the measurements of the photoionisation cross sections of the excited $^1$P$_1$ and $^3$S$_1$ states of ultracold $^{88}$Sr atoms at 389.889 nm wavelength, which is the magic wavelength of the $^{1}$S$_{0}$-${}^{3}$P${}_{0}$ clock transition. The photoionisation cross section of the $^1$P$_1$ state is determined from the measured ionisation rates of $^{88}$Sr in the magneto-optical trap in the $^1$P$_1$ state to be 2.20(50)$\times$10$^{-20}$ m$^2$, while the photoionisation cross section of $^{88}$Sr in the $^3$S$_1$ state is inferred from the photoionisation-induced reduction in the number of atoms transferred through the $^3\text{S}_1$ state in an operating optical lattice clock to be $1.38(66)\times$10$^{-18}$ m$^2$. Furthermore, the resulting limitations of employing a blue-detuned magic wavelength optical lattice in strontium optical lattice clocks are evaluated. We estimated photoionisation induced loss rates of atoms at 389.889 nm wavelength under typical experimental conditions and made several suggestions on how to mitigate these losses. In particular, the large photoionisation induced losses for the $^3$S$_1$ state would make the use of the $^3$S$_1$ state in the optical cycle in a blue-detuned optical lattice unfeasible and would instead require the less commonly used $^3$D$_{1,2}$ states during the detection part of the optical clock cycle.

physics.atom-ph↗

arXiv2022

Distributed distance domination in graphs with no $K_{2,t}$-minor

We prove that a simple distributed algorithm finds a constant approximation of an optimal distance-$k$ dominating set in graphs with no $K_{2,t}$-minor. The algorithm runs in a constant number of rounds. We further show how this procedure can be used to give a distributed algorithm which given $\epsilon>0$ and $k,t\in \mathbb{Z}^+$ finds in a graph $G=(V,E)$ with no $K_{2,t}$-minor a distance-$k$ dominating set of size at most $(1+\epsilon)$ of the optimum. The algorithm runs in $O(\log^*{|V|})$ rounds in the Local model. In particular, both algorithms work in outerplanar graphs.

cs.DS↗

arXiv2020

Adjacency posets of outerplanar graphs

Felsner, Li and Trotter showed that the dimension of the adjacency poset of an outerplanar graph is at most 5, and gave an example of an outerplanar graph whose adjacency poset has dimension 4. We improve their upper bound to 4, which is then best possible.

math.CO↗

arXiv2019

Counterexamples to Thomassen's conjecture on decomposition of cubic graphs

We construct an infinite family of counterexamples to Thomassen's conjecture that the vertices of every 3-connected, cubic graph on at least 8 vertices can be colored blue and red such that the blue subgraph has maximum degree at most 1 and the red subgraph minimum degree at least 1 and contains no path on 4 vertices.

math.CO↗

arXiv2017

Dual Hg-Rb magneto-optical trap

We present a two-species laser cooling apparatus capable of simultaneously collecting Rb and Hg atomic gases into a magneto-optical trap (MOT). The atomic sources, laser system, and vacuum set-up are described. While there is a loss of Rb atoms in the MOT due to photoionization by the Hg cooling laser, we show that it does not prevent simultaneous trapping of Rb and Hg. We also demonstrate interspecies collision-induced losses in the ${}^{87}$Rb-${}^{202}$Hg system.

physics.atom-ph↗

arXiv2014

Hamilton Cycles in Random Lifts of Graphs

For a graph $G$ the random $n$-lift of $G$ is obtained by replacing each of its vertices by a set of $n$ vertices, and joining a pair of sets by a random matching whenever the corresponding vertices of $G$ are adjacent. We show that asymptotically almost surely the random lift of a graph $G$ is hamiltonian, provided $G$ has the minimum degree at least $5$ and contains two disjoint Hamiltonian cycles whose union is not a bipartite graph.

math.CO↗

arXiv2013

Absolute Frequency Measurement of Rubidium 5S-7S Two-Photon Transitions

We report the absolute frequency measurements of rubidium 5S-7S two-photon transitions with a cw laser digitally locked to an atomic transition and referenced to an optical frequency comb. The narrow, two-photon transition, 5S-7S (760 nm) insensitive to first order in a magnetic field, is a promising candidate for frequency reference. The performed tests yield the transition frequency with accuracy better than reported previously.

physics.atom-ph↗

arXiv2013

Matter-Wave Interference versus Spontaneous Pattern Formation in Spinor Bose-Einstein Condensate

We describe effects of matter-wave interference of spinor states in the $^{87}$Rb Bose-Einstein condensate. The components of the F=2 manifold are populated by forced Majorana transitions and then fall freely due to gravity in an applied magnetic field. Weak inhomogeneities of the magnetic field, present in the experiment, impose relative velocities onto different $m_F$ components, which show up as interference patterns upon measurement of atomic density distributions with a Stern-Gerlach imaging method. We show that interference effects may appear in experiments even if gradients of the magnetic field components are eliminated but higher order inhomogeneity is present and the duration of the interaction is long enough. In particular, we show that the resulting matter-wave interference patterns can mimic spontaneous pattern formation in the quantum gas.

cond-mat.quant-gas↗

arXiv2011

Nonrepetitive sequences on arithmetic progressions

A sequence $S=s_{1}s_{2}..._{n}$ is \emph{nonrepetitive} if no two adjacent blocks of $S$ are identical. In 1906 Thue proved that there exist arbitrarily long nonrepetitive sequences over 3-element set of symbols. We study a generalization of nonrepetitive sequences involving arithmetic progressions. We prove that for every $k\geqslant 1$ and every $c\geqslant 1$ there exist arbitrarily long sequences over at most $(1+\frac{1}{c})k+18k^{c/c+1}$ symbols whose subsequences indexed by arithmetic progressions with common differences from the set $\{1,2,...,k\}$ are nonrepetitive. This improves a previous bound obtained in \cite{Grytczuk Rainbow}. Our approach is based on a technique introduced recently in \cite{GrytczukKozikMicek}, which was originally inspired by a constructive proof of the Lovász Local Lemma due to Moser and Tardos \cite{MoserTardos}. We also discuss some related problems that can be successfully attacked by this method.

math.CO↗

arXiv2008

Optimal geometry for efficient loading of an optical dipole trap

One important factor which determines efficiency of loading cold atoms into an optical dipole trap from a magneto-optical trap is the distance between the trap centers. By studying this efficiency for various optical trap depths (2--110 mK) we find that for optimum dipole trap loading, longitudinal displacements up to 15 mm are necessary. An explanation for this observation is presented and compared with other work and a simple analytical formula is derived for the optimum distance between the trap centers.

physics.atom-ph↗