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.

Practical scenarios with multiple simultaneously active speakers recorded using one or more microphones in reverberant rooms pose a challenging problem when the extraction of the desired speaker signal is sought for. The majority of techniques found in the literature facilitate either source separation or dereverberation, which can at best be performed as subsequent, cascade processing. Recently, a solution to the joint task has been proposed, which is known as the weighted power minimization distortionless response (WPD) beamformer. In this paper, we derive a convolutional multichannel filter which performs jointly optimum dereverberation and desired source signal extraction. We formulate a single optimization criterion which minimizes the convolutional source-variance weighted mean square error (CW-MMSE), thereby effectively unifying the weighted prediction error (WPE) based dereverberation and MMSE filtering for the desired source extraction from reverberant mixtures of speakers. Experimental results show a significant performance improvement over the compared state-of-the-art methods such as WPD for datasets with simulated and recorded impulse responses.

We give a distributed algorithm which given ε > 0 finds a (1-ε)-factor approximation of a maximum f-matching in graphs G = (V,E) of sub-logarithmic expansion. Using a similar approach we also give a distributed approximation of a maximum b-matching in the same class of graphs provided the function b: V → ℤ^+ is L-Lipschitz for some constant L. Both algorithms run in O(log^* n) rounds in the LOCAL model, which is optimal.

The recordings of speech in enclosures are corrupted by reverberation caused by multipath wave propagation from the speaker to the distant microphones. In this letter, we address the problem of reducing the late part of room reverberation. The presented blind dereverberation method consists in multichannel linear prediction (MCLP) and enforces sparsity of the dereverberated speech by adopting the split Bregman approach. The proposed algorithm alternately solves two optimization problems, where the former cost function is derived by assuming that speech is modelled using a sparse prior distribution, while the latter optimization emphasizes speech sparsity by an additional incorporation of a weightedl1-norm of the output signal to the standard linear prediction based cost function. The results of experiments performed using simulated and measured room impulse responses for various reverberation time values indicate superior performance of the proposed sparse split Bregman (SSB) method over state-of-the-art non-sparse and sparse MCLP-based dereverberation methods in terms of standard evaluation measures and as pre-processsing to the automatic speech recognition.

In this paper we will give two distributed approximation algorithms (in the Local model) for the minimum dominating set problem. First we will give a distributed algorithm which finds a dominating set D of size O(gamma(G)) in a graph G which has no topological copy of K_h. The algorithm runs L_h rounds where L_h is a constant which depends on h only. This procedure can be used to obtain a distributed algorithm which given epsilon>0 finds in a graph G with no K_h-minor a dominating set D of size at most (1+epsilon)gamma(G). The second algorithm runs in O(log^*{|V(G)|}) rounds.

The VQiPS (Video Quality in Public Safety) Working Group, supported by the U.S. Department of Homeland Security, has been developing a user guide for public safety video applications. According to VQiPS, five parameters have particular importance influencing the ability to achieve a recognition task. They are: usage time-frame, discrimination level, target size, lighting level, and level of motion. These parameters form what are referred to as Generalized Use Classes (GUCs). The aim of our research was to develop algorithms that would automatically assist classification of input sequences into one of the GUCs. Target size and lighting level parameters were approached. The experiment described reveals the experts' ambiguity and hesitation during the manual target size determination process. However, the automatic methods developed for target size classification make it possible to determine GUC parameters with 70 % compliance to the end-users' opinion. Lighting levels of the entire sequence can be classified with an efficiency reaching 93 %. To make the algorithms available for use, a test application has been developed. It is able to process video files and display classification results, the user interface being very simple and requiring only minimal user interaction.