Speech Processing

In this paper we propose the use of an HMM-based phonetic aligner together with a speech-synthesis-based one to improve the accuracy of the global alignment system. We also present a phone duration-independent measure to evaluate the accuracy of the automatic annotation tools. In the second part of the paper we propose and evaluate some new confidence measures for phonetic annotation.

This paper presents the results of our effort in improving the accuracy of a DTW-based automatic phonetic aligner. The adopted model assumes that the phonetic segment sequence is already known and so the goal is only to align the spoken utterance with a reference synthetic signal produced by waveform concatenation without prosodic modifications. Instead of using a single acoustic measure to compute the alignment cost function, our strategy uses a combination of acoustic features depending on the pair of phonetic segment classes being aligned. The results show that this strategy considerably reduces the segment boundary location errors, even when aligning synthetic and natural speech signals of different gender speakers.

Speech is the powerful engine of communication among human beings and language is meant for communicating with the world. This has motivated new researchers to study automatic speech recognition and expand a computer system so it can integrate and understand human speech. But the problem with speech recognition is the acoustic noisy environment can deeply corrupt audio speech. This polluted audio speech disturbs the whole recognition performance. So, the development of AudioVisual Speech Recognition (AVSR) aims to solve the issues by utilizing visual pictures that are undisturbed by noise. This review paper's goal is to explain AVSR architectures, which include front-end operations, the utilized audiovisual dataset, and related studies, audio feature extraction, fusion and modeling techniques, and accuracy estimation methods.

Reconnaissance hors ligne des chiffres Reconnaissance hors ligne des chiffres Reconnaissance hors ligne des chiffres Reconnaissance hors ligne des chiffres manuscrits isolés par manuscrits isolés par manuscrits isolés par manuscrits isolés par l'approche l'approche l'approche l'approche Neuro

In Machine Learning (ML) supervised classification problems, it is often beneficial to crunch down the input
data, mapping it to an initial pre-processing stage, to improve the performance of pattern recognition systems,
such as Artificial Neural Networks (ANNs). But the success of this kind of approach relies on a viable choice
of features to better represent the essence of the input data. In this scenario, most ML sound applications make
use of acoustic models, but, regarding machine hearing, this works better if some basic properties of the human
hearing are considered, such as the variable widths of cochlear critical bands with respect to frequency. In this
work, some techniques of cepstral feature extraction are investigated, based on Linear Predictive Coding (LPC),
which can model the source-filter behavior of glottal speech, and Short-Time Fourier Transform (STFT), using
triangular and gammatone shaped filter banks to warp the short-time power spectrum into different scales of
frequency based on human auditory perception. In addition to it, a dimensional reduction of these representations
is further accomplished by applying a Discrete Cosine Transform (DCT). Subsequently, these feature extraction
techniques are applied to a speech-only portion of an emotion recognition data-set, serving as a front-end to
an ML model. This back-end consists of a shallow and fully connected architecture, known as linear classifier,
which has just one input layer and one output layer, and is regularized via the Least Squares (LS) method. Finally,
each model obtained is compared for accuracy. In validation experiments, the GammaTone Cepstral Coefficients
(GTCC) was the most prominent front-end, with 55% overall accuracy. However, in test experiments, the Mel
Frequency Cepstral Coefficients (MFCC), Linear Predictive Cepstral Coefficients (LPCC) and Bark Frequency
Cepstral Coefficients (BFCC) features resulted in improved accuracy, with 98%, 94% and 94%, respectively.

This essay explores the motif of "the Pit" as a third, thematic focus for the Psalter (alongside "refuge" and "pathway"). Especially as it comes to fullest expression in Psalm 88, the Pit can be seen as the "black hole" at the center of the Psalms. Yet, while powerful, Psalm 88, and the presence of the Pit therein, does not have the final word. The sequence of the following psalms (Psalms 89, 90, and 91) slowly but surely draws the poet of Psalm 88 out of what has been called the saddest of all psalms. In this way, Psalm 88 retains its witness within the Psalter, but it is also not alone within the Psalms. Astrophysics has determined that black holes are not only destructive; they are also generative, shaping things around them and producing massive amounts of energy. In this light, the Pit, too, can be viewed as a generative theme for the Psalms, one that is crucial-somehow even central-but, in the end, transcended.

There have been wide and fundamental changes in the field of speech and language neuropsychology since the publication of Paul Broca’s (1824–1880) epoch-making work on “aphasie” (siège du langage articulé) in 1863 (Broca, 1863). This Research Topic surveys the efforts to understand the relationship between human behavior and brain function with respect to language, cognition, and memory with a focus on activities from the 1860s to 1960s in Europe and North America. The reviewed period begins with the groundbreaking work of Broca in France, John Hughlings Jackson (1835–1911) in Great Britain, and Carl Wernicke (1848–1905) in Germany to identify the neuropathological sources of selective impairments in language (Figure 1) (Levelt, 2013).
    Efforts continued throughout the second half of the 19th century, leading to increased activity in the wake of the two World Wars. One hundred years later, interest resurged in the earlier ideas when new approaches were initiated by individuals such as Wilder Penfield (1891–1976) in Canada (Penfield, 1949) and Norman Geschwind (1926–1984) in the United States (Geschwind, 1970). Historiographical approaches to the understanding
of these efforts have considered the applied models and metaphors of speech and language neuropsychology, methodological approaches in the clinic and laboratory, as well as on the status of evidence, the flow of ideas and people, along with interdisciplinary exchanges with anthropology, education, linguistics, medicine, and sociocultural contexts (Eling, 1994).
The scholarly collaboration showcased in this Research Topic has generated novel insights and stimulating perspectives on the emergence of speech and language neuropsychology. For instance, the twenty contributing authors have investigated phases and events of the long-standing debate between localizationists and holists in the field of neuropsychology. They have analyzed how and why new concepts and theories have emerged, including for clinical and rehabilitation purposes (Stahnisch and Hoffmann, 2010). Furthermore, the limits imposed by certain models on basic and clinical research since Broca’s and Wernicke’s times were investigated (Tremblay and Dick, 2016).

Speech summarization has become an essential tool for efficiently managing and accessing the growing volume of spoken and audiovisual content. However, despite its increasing importance, speech summarization is still not clearly defined and intersects with several research areas, including speech recognition, text summarization, and specific applications like meeting summarization. This survey not only examines existing datasets and evaluation methodologies, which are crucial for assessing the effectiveness of summarization approaches but also synthesizes recent developments in the field, highlighting the shift from traditional systems to advanced models like fine-tuned cascaded architectures and end-to-end solutions.

The accuracy and computational complexity of keyword spotting (KWS) systems are heavily influenced by the choice of audio features in speech signals. This paper introduces a novel approach for audio feature extraction in KWS by leveraging a convolutional autoencoder, which has not been explored in the existing literature. Strengths of the proposed approach are in the ability to automate the extraction of the audio features, keep its computational complexity low, and allow accuracy values of the overall KWS systems comparable with the state of the art. To evaluate the effectiveness of our proposal, we compared it with the widely-used Mel Frequency Cepstrum (MFC) method in terms of classification metrics in noisy conditions and the number of required operators, using the public Google speech command dataset. Results demonstrate that the proposed audio feature extractor achieves an average classification accuracy on 12 classes ranging from 81.84% to 90.36% when the signal-tonoise ratio spans from 0 to 40 dB, outperforming the MFC up to 5.2%. Furthermore, the required number of operations is one order of magnitude lower than that of the MFC, resulting in a reduction in computational complexity and processing time, which makes it well-suited for integration with KWS systems in resource-constrained edge devices.

Functional MRI was used to investigate the characteristics of the human cerebral response to dynamic ripples. Dynamic ripples are sound stimuli containing regular spectrotemporal modulations, which are of major importance in speech processing; however, in contrast to speech, dynamic ripples can be characterized fully by a limited number of parameters. Extensive activation consisting of multiple separate regions was found bilaterally in the auditory cortex, particularly along the Heschl's gyri. This agrees with the presence of a structural cortical subdivision into functional fields. The level and the extent of activation were measured and correlated highly (R 2 ϭ 0.97). Both measures depended strongly on the spectral density, temporal frequency, and amplitude of the modulations and matched the perceptual discernibility of the spectrotemporal modulations. The largest responses occurred for parameter values near the optimal human sensitivity. The drift direction of the modulations did not influence the activation. No quantitative differences were found between the two hemispheres. Average brain activation levels proved to be separable with regard to the spectral density and temporal frequency of the modulations. Topographic mappings of the modulation density and frequency onto the cortical surface were shown, approximately in posterolateral-to-anteromedial and lateral-to-medial directions, respectively. Posterolateral regions were most sensitive to spectrotemporal features at a scale similar to phonemes. Anteromedial regions, however, were also relatively sensitive to smaller scale acoustic features. This spatially dependent sensitivity suggests a functional topographic and hierarchical organization of the auditory cortex.

This  study  identifies,  categorizes,  and  analyzes speech  errors  committed  by  EFL  students  and explores  their  implications  for  English  teaching. Data  were  collected  through  contentanalysis, following  a  qualitative  descriptive  design,  and recorded  via  an  audio  recorder  and  a  speech  log over  one  month.  The  study  revealed  that  speech errors    were    categorized    into    twelve    types: anticipation, perseveration, substitution, metathesis  (exchange),  addition,  blends,  shift, omission  (deletion),  malapropism,  spoonerism, sound-exchange  errors,  and  native  language interference errors. To improve students' speaking skills,  English  teachers  may  design  lessons  that enhance fluency and accuracy while encouraging EFL  learners  to  develop  self-repair  strategies  to minimize  repeated  errors.  Additionally,  they  are encouraged    to    conduct    further    research    on effective  error  correction  strategies  to  improve teaching practices.

This review examines how researchers have approached speech errors across a wide range of studies, drawing on work from databases such as Scopus, Web of Science, and Google Scholar. Rather than beginning with fixed categories, patterns were allowed to emerge gradually through close engagement with the literature. The analysis brought to light six recurring themes: (1) speech errors are explainable, often reflecting identifiable processes; (2) they are traceable and identifiable across different contexts; (3) they are largely unavoidable, yet speakers frequently manage to repair them in interaction; (4) they serve as valuable evidence for understanding how language functions in real-time; (5) they are consistently interpreted through established theoretical frameworks; and (6) they have remained a subject of continued interest, with recent studies revisiting and extending earlier insights. Taken together, these themes point to speech errors as systematic, shaped by cognitive processes, and far from random or trivial. Beyond showing how language is processed and produced, they offer practical insights, particularly for English Language Teaching (ELT). At the same time, the review highlights areas where further research is needed, especially on lesser-studied languages, multilingual contexts, and the potential of emerging technologies for speech analysis. By bringing these strands together, the review offers a more connected view of speech errors and their relevance to language education, psycholinguistics, and communication studies.

Transcription systems are crucial for linguistic research, language preservation, and computational applications. However, existing systems, such as the International Phonetic Alphabet (IPA), often lack phonemic specificity and require extensive learning. NAVLIPI, a novel system introduced by Prasanna Chandrasekhar in 2012, presents a comprehensive, intuitive, and universal alphabet capable of representing all the world's languages. Built upon the Latin script and supplemented with additional symbols, NAVLIPI addresses key phonemic distinctions while maintaining phonetic precision. This paper explores the structure of NAVLIPI, its advantages over existing transcription systems, and its potential applications in linguistics, education, and natural language processing.

This paper presents a method for extracting MFCC parameters from a normalised power spectrum density. The underlined spectral normalisation method is based on the fact that the speech regions with less energy need more robustness, since in these regions the noise is more dominant, thus the speech is more corrupted. Less energy speech regions contain usually sounds of unvoiced nature where are included nearly half of the consonants, and are by nature the least reliable ones due to the effective noise presence even when the speech is acquired under controlled conditions. This spectral normalisation was tested under additive artificial white noise in an Isolated Speech Recogniser and showed very promising results . It is well known that concerned to speech representation, MFCC parameters appear to be more effective than power spectrum based features. This paper shows how the cepstral speech representation can take advantage of the abovereferred spectral normalisation and shows some results in the continuous speech recognition paradigm in clean and artificial noise conditions.

Normally, voice activity detection (VAD) refers to speech processing algorithms for detecting the presence or absence of human speech in segments of audio signals. In this paper, however, we focus on speech detection algorithms that take VoIP traffic instead of audio signals as input. We call this category of algorithms network-level VAD. Traditional VAD usually plays a fundamental role in speech processing systems because of its ability to delimit speech segments. Network-level VAD, on the other hand, can be quite helpful in network management, which is the motivation for our study. We propose the first real-time network-level VAD algorithm that can extract voice activity from encrypted and non-silence-suppressed Skype traffic. We evaluate the speech detection accuracy of the proposed algorithm with extensive reallife traces. The results show that our scheme achieve reasonably good performance even high degree of randomness has been injected into the network traffic.

This paper proposes an efficient method based on the steered-response power (SRP) technique for sound source localization using microphone arrays: the volumetric SRP (V-SRP). As compared to the SRP, by deploying a sparser volumetric grid, the V-SRP achieves a significant reduction of the computational complexity without sacrificing the accuracy of the location estimates. By appending a fine search step to the V-SRP, its refined version (RV-SRP) improves on the compromise between complexity and accuracy. Experiments conducted in both simulated-and real-data scenarios demonstrate the benefits of the proposed approaches. Specifically, the RV-SRP is shown to outperform the SRP in accuracy at a computational cost of about ten times lower.

Recent developments in large vocabulary continuous speech recognition (LVCSR) have shown the effectiveness of discriminative training approaches, employing the following three representative techniques: discriminative Gaussian training using the minimum phone error (MPE) criterion, discriminately trained features estimated by multilayer perceptrons (MLPs); and discriminative feature transforms such as feature-level MPE (fMPE). Although MLP features, MPE models, and fMPE transforms have each been shown to improve recognition accuracy, no previous work has applied all three in a single LVCSR system. This paper uses a state-of-the-art Mandarin recognition system as a platform to study the interaction of all three techniques. Experiments in the broadcast news and broadcast conversation domains show that the contribution of each technique is nonredundant, and that the full combination yields the best performance and has good domain generalization.

A lower-error and lower-variance n X ?Z multiplier is suitably proposed for VLSI design. Considering next lower significant stage in P,-' column and useful error-compensation model in the least significant part, and utilizing a near optimized index to classify the error terms are our strategies in order to achieve lower error and variance as compared with previously proposed structure in the subproduct-array of Baugh-Wooley algorithm. This novel structure applied to the fixed-width lowpass digital FIR filter for speech signal processing system has excellent performance in reducing maximum error, average error, and variation of errors as shown in given tables and figures.

One of the practical issues of the voice activity detection (VAD) algorithms, that are commonly installed in today's call centres, lies in the tremendous variability of call recording conditions. Particularly, the intermittent occurence of noise-only or silent speech recordings among normal volume recordings of standard talks is very displeasing and often significantly decreases the overal true detection rates. We describe an enhancement of statistical VAD algorithm, which makes it more robust to most of such deviations, while not exalting its overal complexity.

Given the importance of speech in our day-to-day activities and almost all forms of communication, it is crucial to address issues that hinder effective communication. Often, in critical security and military applications, the intelligibility of speech is of higher importance than the quality. In such cases, it is crucial that the speech utterance is comprehended for the exact message it was meant to convey, while the pleasantness of the speech and how good it sounds assumes secondary importance. In cases where noise degrades speech intelligibility, it is vital to develop measures that allow speech utterances to retain their intelligibility despite the ambient noise. When modified in a certain way, the intelligibility is improved over the original voice in the presence of noise. This is known as the Lombard effect. Of the several techniques to achieve Lombard speech, one effective approach is formant shifting, where the formant frequencies are relocated to regions in the spectrum where the noise cannot mask the phonetic information they contain. This shift was guided by a trapezoidal voice transformation function (TVTF) that mapped the original locations of formant frequencies in the spectrum to new locations. However, when done empirically, formant shifting results in artifacts.

{Hence, the objective of this thesis is to ascertain techniques to optimize such shifts in formants in order to maximize the boost in intelligibility in a near-end noisy environment. Such optimization should encompass factors like the language being spoken, the presence of realistic noises that are frequently encountered, as well as variation in noise intensities in terms of signal SNR.} {To that end, as the first contribution of this thesis}, we propose optimizing the shaping parameters of the TVTF via a genetic optimization technique called comprehensive learning particle swarm optimization (CLPSO). {Such optimization was specific to a certain combination of language, noise type, and SNR level. Next}, we propose a joint enhancement scheme where we combine several voice modification techniques with formant shifting to preserve speech quality while boosting intelligibility. These include time-scale modification, energy redistribution, and smoothing of formant contours. Although the performance increased, the optimization, which was already computationally intensive, took even longer to converge due to the increased dimension of the input vector corresponding to the additional techniques used in Lombard speech generation. To address this, {as yet another contribution,} a statistical configuration of the VTF is proposed - the Gaussian voice transformation function (GVTF), which had just three parameters to be optimized (instead of five in the TVTF). This reduced the convergence time and resulted in a smoother contour of the shifted formants across the frames.

In case of changes in the ambient conditions of SNR levels, language, and/or noise type, {our contributions focus on the proposition of} transfer learning on all environmental factors like SNR, languages, and noise types. The transfer learning across SNR levels was achieved through a Gaussian process regression, where the known VTF parameters at some SNR levels were used to estimate the unknown parameters for other SNR levels. In case of a change in the language, the comparative analysis of pitch and formant frequencies between the source and target languages was used to modify the shaping parameters of the VTF to conform to the target language. However, the transfer across noises was only made possible through GVTF. This is handled through comparative analysis of the Gaussian approximations of the noise magnitude spectra of the source and target noises. Finally, we establish a mechanism for combined transfer learning across languages and noises for dealing with cases where both those conditions change.

To demonstrate the efficacy of the proposed algorithms in real-time, the entire optimization cycle, using TVTF and GVTF, along with all transfer learning mechanisms, have all been built into an application user interface developed in the AppDesigner in MATLAB. As a practical application of the optimization used in this thesis, the importance of an optimal dataset generated through CLPSO has been demonstrated towards training a lightweight ANN on a resource-constrained Raspberry Pi to suggest optimal microphone location in a room relative to a speaker.

A novel technique based on dynamic stochastic resonance (DSR) in discrete cosine transform (DCT) domain has been proposed in this paper for the enhancement of dark as well as low-contrast images. In conventional DSR-based techniques, the performance of a system can be improved by addition of external noise. However, in the proposed DSR-based work, the intrinsic noise of an image has been utilized to create a noise-induced transition of a dark image to a state of good contrast. The proposed technique significantly enhances the image contrast and color information without losing any image or color data by optimization of bistable system parameters. The performance of the proposed methodology has been measured in terms of relative contrast enhancement factor, perceptual quality measure, and color enhancement factor. When compared with the existing enhancement techniques, such as adaptive histogram equalization, gamma correction, single-scale retinex, multi-scale retinex, modified high-...

In many functional magnetic resonance imaging (fMRI) studies blind humans were found to show cross-modal reorganization engaging the visual system in non-visual tasks. For example, blind people can manage to understand (synthetic) spoken language at very high speaking rates up to ca. 20 syllables/s (syl/s). FMRI data showed that hemodynamic activation within right-hemispheric primary visual cortex (V1), bilateral pulvinar (Pv), and left-hemispheric supplementary motor area (pre-SMA) covaried with their capability of ultra-fast speech (16 syllables/s) comprehension. It has been suggested that right V1 plays an important role with respect to the perception of ultra-fast speech features, particularly the detection of syllable onsets. Furthermore, left pre-SMA seems to be an interface between these syllabic representations and the frontal speech processing and working memory network. So far, little is known about the networks linking V1 to Pv, auditory cortex (A1), and (mesio-) frontal areas. Dynamic causal modeling (DCM) was applied to investigate (i) the input structure from A1 and Pv toward right V1 and (ii) output from right V1 and A1 to left pre-SMA. As concerns the input Pv was significantly connected to V1, in addition to A1, in blind participants, but not in sighted controls. Regarding the output V1 was significantly connected to pre-SMA in blind individuals, and the strength of V1-SMA connectivity correlated with the performance of ultra-fast speech comprehension. By contrast, in sighted controls, not understanding ultra-fast speech, pre-SMA did neither receive input from A1 nor V1. Taken together, right V1 might facilitate the "parsing" of the ultra-fast speech stream in blind subjects by receiving subcortical auditory input via the Pv (= secondary visual pathway) and transmitting this information toward contralateral pre-SMA.

In recent years, deep learning approaches have gained significant interest as a way of building hierarchical representations from unlabeled data. Additionally, in the field of sound direction-of-arrival (DOA) estimation, the binaural features like interaural time or phase difference and interaural level difference, or monaural cues like spectral peaks and notches are often used to estimate sound DOA. Although these binaural or monaural cues successfully explained human sound DOA ability, its accuracy and extent are all limited to the human knowledge on feature extracting methods. In this paper, we are interested in applying a deep learning approach to monaural sound localization based on monaural spectral cues. A convolutional deep belief network is applied to monaural auditory spectrograms processed by a computational auditory model to learning monaural features automatically. The learned features are then regressed using a support vector regression (SVR) model for sound DOA estimation tasks. Moreover, our feature representations learned from unlabeled monaural auditory spectrograms are then compared to the well-known binaural features. The results indicate that our monaural model shows reasonable performance at the task of 3D DOA estimation.

This paper proposes an acoustic phonetic study of the foreign accents in the Arabic language. To analyze on a large scale of the connected variations, the contribution of the automatic tools acoustico-phonetic decoding tools along the adaptation forces. Based on the results of the uses of these tools, we managed to differentiate between three major forms of accents: The English language speaker accent (American and British population's accent), the European people accent (French speaker accent, German speaker accent) and the Persian speaker accent.

In this paper, we propose an emotion recognition system from speech signal using both spectral and prosodic features. Most traditional systems have focused on spectral features or prosodic features. Since both the spectral and the prosodic features contain emotion information, it is believed that combining spectral features and prosodic features will improve the performance of the emotion recognition system. Therefore, we propose to use both spectral and prosodic features. For spectral features, a GMM super vector based SVM is applied. For prosodic features, a set of prosodic features that are clearly correlated with speech emotional states and SVM is also used for emotion recognition. The combination of both spectral features and prosodic features is done and an SVM is trained using the combined feature vector. The emotion recognition accuracy of our experiments allow us to explain which features carry the most emotional information and why. It also allows us to develop criteria to...

Several techniques have been introduced over the years to mitigate the effect of howling sound production in sound reinforcement applications. The algorithmic computational complexity of the howling controllers directly affects their implementability on resource-constrained hardware platforms. This work introduces a computationally cheap acoustic howling control algorithm that requires a single multiplication operation for its implementation. The algorithm was modelled, simulated and tested using MATLAB® to obtain its performance metrics. Simulation results showed that the developed technique has average sound quality indices of 1.1746 and 1.2118 for speech and audio signals respectively. Also, for speech and audio signals, Maximum Stable Gain (MSG) Increase figures of 11.29 dB and 9.858 dB were obtained respectively. These results showed that, compared to the standard Adaptive Feedback Control (AFC) technique, the developed algorithm has comparable MSG Increase figures, negligibly small computational cost and low sound quality. This result implied that the howling controller can be easily implemented on a resource-constrained hardware platform at the expense of some degradation in sound quality.

1 It has been observed that in a HLH (High-Low-High) tone sequence, the second H tone is lowered in pitch, as compared to a HHH tone sequence, which was termed as downstep. To calculate the downstep effect and test its scope, we compared sentences in the tonal structure of (HHHHH H )(HHHHHHH) to (HHHHH L )(HHHHHHH). The syntactic boundary between the L (X) and the H (Y) was either a syllable boundary or a phrase boundary. Moreover, corrective focus was elicited either on the syllable X, Y, sentence final (Z) or none (wide focus). The results showed that: (1) in the wide focus and sentence-final focus (ZF) condition, downstep can last for several syllables; (2) a phrase boundary did not block downstep; (3) when syllable Y was focused, on-focus F0 raising did not override downstep, however, it stopped the trend of downstep in the following H tones; (4) when syllable X was focused, it caused a post-low-bouncing effect on the following H tones instead. Taken together the data showed tha...

У статті наведено теоретичне узагальнення й нове розв’язання наукової проблеми формування термінологічного апарату ґрунтознавства в процесі наукового аналізу та підготовки виконання про- фесійних завдань майбутніх фахівців. Визначено методичні засади формування термінологічного апа- рату в галузі ґрунтознавства. Проаналізовано наукові дослідження та літературні джерела щодо методів і підходів до вивчення термінології ґрунтознавства.
Висвітлено роль наукового аналізу в управлінні та оптимізації процесу формування термінологічного апарату, а саме: стандартизації термінології, що дає змогу визначити ключові терміни та забезпечує однозначне розуміння понять, їх визначення, полегшує спілкування між фахівцями; адаптації до
нових відкриттів та методів, що може призвести до оновлення або розширення термінологічного
апарату; уникнення синонімізму та плутанини, які можуть виникати через використання різних тер- мінів для позначення одних і тих самих понять або через нечіткі визначення; забезпечення точності й
зрозумілості; оптимізований термінологічний апарат може сприяти розвитку ґрунтознавства, зрос- танню інтересу до досліджень у цій сфері та впровадженню нових підходів у практику. Доведено, що формування термінологічного апарату в курсі Ґрунтознавства є системою, інтегра- тивною єдністю таких складників, як: вивчення базових понять, розуміння взаємозв’язків, практичні
навички, стале поповнення знань, застосування міждисциплінарних знань.
Обґрунтовано взаємозв’язок між формуванням термінологічного апарату та успішним виконанням
професійних завдань майбутніми фахівцями у сфері ґрунтознавства, агрономії, геодезії та землеу- строї. Акцентовано, що термінологічний апарат ґрунтознавства допомагає студентам розробляти
комплексні стратегії в галузі агрономії та геодезії, що відповідають вимогам сучасного ринку праці
й сприяє їхній успішній професійній інтеграції після завершення навчання, забезпечуючи необхідні
інструменти для ефективної роботи, дає змогу майбутнім фахівцям чітко спілкуватися з колегами,
розуміти наукову літературу та виконувати різноманітні аналітичні завдання, визначати проблеми в
ґрунтознавстві й розробляти ефективні рішення для їх вирішення, чіткість та уніфікація термінології
сприяє точності й надійності результатів досліджень, підвищення професійного статусу.

This article presents a full end-to-end pipeline for Arabic Dialect Identification (ADI) using intonation patterns and acoustic representations. Recent approaches to language and dialect identification use linguistic-aware deep architectures that are able to capture phonetic differences amongst languages and dialects. Specifically, in ADI tasks, different combinations of linguistic features and acoustic representations have been successful with deep learning models. The approach presented in this article uses intonation patterns and hybrid residual and bidirectional LSTM networks to learn acoustic embeddings with no additional linguistic information. Results of the experiments show that intonation patterns for Arabic dialects provide sufficient information to achieve state-of-the-art results on the Var-Dial 17 ADI dataset, outperforming single-feature systems. The pipeline presented is robust to data sparsity, in contrast to other deep learning approaches that require large quantities of data. We conjecture on the importance of sufficient information as a criterion for optimality in a deep learning ADI task, and more generally, its application to acoustic modeling problems. Small intonation patterns, when sufficient in an information-theoretic sense, allow deep learning architectures to learn more accurate speech representations.

The autonomous vehicle market is experiencing significant growth, with indications of transitioning from the "trough of disillusionment" to the "slope of enlightenment" on the Gartner hype cycle chart. Fundamental technologies encompassing extensive data analytics, computational capabilities, and sensor fusion techniques have already been established, and all stakeholders in this industry are persistently exploring novel approaches to enhance the overall perception of end users in terms of safety and trustworthiness. In this context, this project aims to develop and implement an Environmental Sound Recognition (ESR) algorithm in an embedded system for deployment in autonomous vehicles for Smart Cities in 2025, targeting advanced functionalities for early warning systems. Due to hardware constraints, a regular passenger vehicle was used, embedding the ESR algorithm in a Raspberry Pi with a microphone array. The limited literature on ESR algorithms for vehicles primarily focuses on siren detection without real-time inferences, and to address this, a dataset benchmarking study confirmed classifiers’ accuracy, leading to the creation of a new dataset tailored to autonomous vehicles. This new dataset provided a comprehensive baseline where several classifiers were trained and evaluated for accuracy, memory usage, and prediction time, with CNN 2D using aggregated features emerging as the top-performing model, achieving an average accuracy of 80% in the sliding window process. During the indoor experiment, the total prediction time attained an average of 47.6 ms, validating the algorithm’s performance with weighted F1-scores close to or better than cross-validation results. In the final phase of the methodology, real-world tests conducted in a passenger vehicle yielded similar results. However, inconsistencies were observed in certain classes due to insufficient sample diversity and environmental noise, which affected their accuracy. The results of this project indicate that its general objective was successfully achieved, contributing to understanding of ESR algorithms in embedded systems within passenger vehicles, and it is ready for integration into the electric and electronic architecture of autonomous vehicles for Smart Cities. Additionally, upon conducting further experiments across various vehicle categories to assess cabin insulation effects, this project could potentially enhance safety features for drivers with hearing impairments by adapting the ESR algorithm as an add-on feature in regular passenger vehicles.

This paper investigates the performance of a selection of state-of-the-art array signal-processing techniques for the purpose of predicting the binaural listening experiments from the equalization and cancellation (EC) paper by Durlach written in 1963. Two different array signal-processing techniques are analyzed, 1) filter and sum beamforming (FS), and 2) minimum variance distortionless response (MVDR) beamforming. The theoretical properties of these beamformers for the specific situation of prediction of binaural masking level differences are analyzed in conjunction with the EC model. Also, the performance of the different beamformers on the data sets in the Durlach paper from 1963 is compared to the EC model.

This study delves into designing intelligent voice assistants through the implementation of opensource speech recognition algorithms. Developers can build AI-powered voice interfaces by utilizing technologies such as Whisper, DeepSpeech, and Kaldi, enabling them to process spoken language, understand user goals, and produce responses. Open-source options provide flexibility and cost-effectiveness, but difficulties persist in terms of accuracy disparities, background noise disturbances, and the computational requirements that accompany them. Protecting user privacy and preventing unauthorized access requires secure data handling practices. Improvements in deep learning, multilingual capabilities, and bespoke customization can improve speech recognition performance and user interaction. Implementing emotion-sensing artificial intelligence and contextual memory capabilities can enhance conversational flow and responsiveness. Responsible AI deployment necessitates the resolution of ethical concerns. Future studies should concentrate on refining AI models to improve real-time processing capabilities and enhance their comprehension of context. Continuous enhancements enable open-source voice assistants to serve as a viable alternative to proprietary systems. This research focuses on important methods, obstacles, and possible improvements in the creation of voice assistants.

In this paper a methodology for extraction of features in emotional speech recognition is presented. Different emotional states of a speaker produce physiological changes in the human speech system, which is reflected in the variation of voice parameters. Processing techniques applied in this work are: time-frequency representations, linear prediction and raw data.

Employing the methodology of Editorial Criticism, this article seeks to demonstrate that Book II of the Psalter (i.e., Psalms 42-72) consists of three parallel, compositional arcs that take the form of a journey. Based on keyword links, as well as clear organizational clues found in the superscriptions, the psalms of Book II are deliberately arranged to portray three parallel depictions of the same progression of events. Mirroring the general sequence of eschatological events portrayed in the prophets, Book II presents a threefold analogous deliverance out of exile into a global kingdom ruled by the Davidic king. While each journey highlights one compelling facet of that future kingdom, together they provide a multiperspectival and even complementary vision of the future which signals a renewed hope in the Davidic promises. Thus, this article will argue that the design of Book II of the Psalter does not extol the glorious days of a bygone era, but rather orients the reader to anticipate the ultimate eschatological fulfillment of what God had guaranteed in the future. The net effect of these compositional arcs (or journeys) for the readers would inevitably be to sustain them through the disorienting experiences of captivity and provide invincible hope for the age to come.

En 1953, C. Baur édite trois homélies festives pseudo-chrysostomiennes, qui, d'après lui, proviendraient de l'entourage de Nestorius sinon de celuici même 1 . L'une s'intitule Sur la sainte Pâque 2 , les deux autres Sur l'Ascension de Notre Seigneur Jésus-Christ 3 . Selon lui, la pensée et l'expression reflètent les discussions christologiques qui ont commencé entre Eusthate d'Antioche et Diodore de Tarse, qui se sont poursuivies avec Théodore de Mopsueste et qui ont atteint un point culminant avec Nestorius. Mais l'attribution de ces textes à Chrysostome serait due aux circonstances historiques et ecclésiales: Théodose avait ordonné de brûler les ouvrages de Nestorius à la suite de sa condamnation 4 . La tradition manuscrite atteste qu'assez tôt les oeuvres de Nestorius ont été transmises sous le patronage d'un nom célèbre à Antioche, Jean Chrysostome, pour échapper à leur destruction 5 . Ces trois homélies pourraient appartenir, selon C. Baur, à cette catégorie de textes. Toutefois, les deux homélies sur l'Ascension sont très différentes tant par leur coloration scripturaire que par les caractéristiques de la fête qu'elles mettent en avant, l'une étant marquée par un trait de polémique anti-juive à la différence de l'autre. On peut légitimement penser qu'elles ont donc été écrites dans des circonstances ou pour des destinataires bien distincts. Telle est la raison pour laquelle nous avons pris le parti de ne pas les examiner ensemble et de présenter une étude de l'homélie CPG 4739 pour elle-même 6 . C. Baur a édité le texte à partir d'un choix de douze manuscrits, après en avoir consulté dix-sept, sans ignorer en outre l'existence du Messina S. Salvatore 3 (XII s.) et du manuscrit d'Istanbul, Megalê tou genous Scholê, 62 (1) daté de 1373 7 . Le texte figure dans des collections d'homélies liturgiques en l'honneur des fêtes du Christ voire de Marie, au milieu de laquelle se trouve la plupart du temps, un sous-corpus sur l'Ascension 8 . L'homélie CPG 4739 est le plus souvent associée aux homélies de Basile de Séleucie (CPG 6659) 9 et du Ps.-Eusèbe d'Alexandrie (CPG 5528) 10 , et dans une moindre mesure à une autre homélie pseudo-chrysostomienne (CPG 4533) 11 ainsi qu'à l'une de Chrysostome lui-même (CPG 4342) 12 . Nous reproduisons ci-dessous le texte édité par C. Baur, accompagné du titre par lequel la tradition manuscrite a transmis l'homélie sous le nom de Jean Chrysostome, en y adjoignant une traduction. 6. La tradition manuscrite a très rarement associé les deux textes. D'après nos enquêtes sur la base de données Pinakes (), seul le manuscrit Berlin gr. 77, Phillipps 1481 (XII e s.) présente les deux homélies l'une à la suite de l'autre. 7. La base Pinakes recense d'autres manuscrits non mentionnés pas C. Baur: le Meteora 138 (XVI e s.); le Messina S. Salv. 092 (XI e s.), Bibliotheca Regionale Universitaria, folios 174-187; le Messina S. Salv. 026 (XII e -XIII e s.), folios 43v-45v; l'Athena 0282 (XVIe s.), EBE, folios 275-277v; Oxford, Holkham gr. 022 (XV e -XVI e s.), Bodl. Libr., folios 134v-136v; Paris, gr. 1186 (XIV e s.), BnF, folios 200-201v.; Tyrnabo 33 (XVI e s.), Dêmotikê bibl., folios 312v-315. En revanche, plusieurs manuscrits utilisés par C. Baur ne se trouvent pas répertoriés par Pinakes: Cod. Ottobon. gr. 411 (XIV e -XV e s.); Vat. gr. 1190 (XVI e s.), Athos, Monè Batopèdiou, 633 (1422); Venedig, Marcian. gr. 43 (XV e s.); Vaticanus gr. 564 (XII e s.); Paris gr. 1175 (XI e s.); Saloniki, monè tôn Blataïôn 6 (X e s.). 8. Nous renvoyons au descriptif sommaire de chaque manuscrit présenté par C. Baur en introduction de son édition, Drei unedierte Festpredigten aus der Zeit der nestorianischen Streitigkeiten, pp. 111-112, ainsi qu'à la base Pinakes pour le détail de chaque manuscrit mentionné. 9. D'après les indications de la base Pinakes, elle figure dans neuf manuscrits avec l'homélie sur l'Ascension de Basile de Séleucie. 10. D'après le même outil, on la trouve dans douze manuscrits avec le texte du Ps. Eusèbe d'Alexandrie. 11. Les deux textes figurent ensemble dans sept manuscrits. 12. On trouve les deux homélies associées dans sept manuscrits.

This paper documents a submission to the shared task on scene segmentation hosted at KONVENS 2021 (Zehe et al., 2021b). The aim of this shared task was to find methods for segmenting narrative texts into different scenes -segments of text where location, time and the constellation of characters stay more or less coherent. This task is formulated as a sentence classification task where sentences bordering the scenes have to be distinguished from in-scene sentences. The approach presented in this paper is based on two steps. In the first one, a twin BERT training setup is used to learn a sentence embedding space in which sentences functioning as scene borders are well-separated from ones that are in-scene. In the second one, the sentence embeddings generated by this model are used as feature vectors to feed a gradient-boosted decision tree ensemble which conducts final predictions. In the shared task leaderboard, the system ranked second in track 1 and first in track 2.

In numerous applications that rely on audible voice signals, including speech recognition, audio recording, and telecommunications, the suppression of background noise is an essential component. The present study introduces an innovative methodology for mitigating noise in audio transmission by employing digital signal processing. Preceding post-processing for refinement, it is necessary to estimate and reduce noise in the input speech signals. The successful implementation of the proposed approach requires this. The proposed method incorporates spectral shaping as a post-processing step. Noise estimation is performed using minimal statistics, and spectrum analysis is executed via short-time Fourier transform. The studies' findings indicate that the intelligibility and audibility of communication in chaotic environments are significantly improved by the proposed method. Anthology-perceptual evaluation scores of voice quality, improved signalto-noise ratios, and decreased word error rates are some objective criteria utilized in quantitative evaluation. Furthermore, subjective hearing tests were employed to authenticate the exceptional quality of the speech signals that were processed. The method that was suggested has demonstrated encouraging outcomes, suggesting that it might be viable to improve the functionality of voice-reliant applications in real-world scenarios. The algorithm implementation is proposed for FPGA and DSP component specific UVM TB verification technique is modelled.

When the audio and visual portions of a speech stimulus are presented synchronously, the resulting enhancement in intelligibility is generally much larger than the one obtained when the audio and visual stimuli are presented sequentially. However, perfect synchronization is not required to obtain a substantial audiovisual (AV) benefit: many studies have shown that AV integration is maximum when the audio signal is slightly delayed relative to the visual signal, and other studies have shown that substantial AV ...

In any signal noise is an undesired quantity, however most of thetime every signal get mixed with noise at different levels of theirprocessing and application, due to which the information containedby the signal gets distorted and makes the whole signal redundant.A speech signal is very prominent with acoustical noises like bubblenoise, car noise, street noise etc. So for removing the noises researchershave developed various techniques which are called filtering. Basicallyall the filtering techniques are not suitable for every application,hence based on the type of application some techniques are betterthan the others. Broadly, the filtering techniques can be classifiedinto two categories i.e. linear filtering and non-linear filtering.In this paper a study is presented on some of the filtering techniqueswhich are based on linear and nonlinear approaches. These techniquesincludes different adaptive filtering based on algorithm like LMS,NLMS and RLS etc., Kalman filter, ARMA and NARMA...

Speech is a form of communication that most people came across in their day to day life. Speech can be used for many purposes like speech communication, speech recognition, speaker identification etc. In all of these applications a noise free speech signal is highly desired. Therefore to get a noise free speech, different speech processing and speech filtering techniques are used. In this paper we have presented a review on speech processing and different speech filtering techniques. The techniques that we have investigated in this paper are Adaptive filtering, Kalman Filtering, Neural Networks and adaptive neuro-fuzzy filtering and it is found that LMS gives good result in linear noise depending upon the order of the filter and hybrid filtering techniques like ANFIS because it gives the better result in case of non-stationary noise. Keywords: Adaptive Filter, noise cancellation, least mean square adaptive filter, Autoregressive, Kalman filter, neural network, Adaptive neuro-fuzzy f...

Nous décrivons un modèle stochastique de type PHMM-Pseudo 2D Hidden Markov Model-, pour la reconnaissance globale de chaînes de caractères arabes imprimés. Le modèle est appliqué directement sur l'image sans segmentation au préalable. La notion de durée est utilisée à la fois horizontalement et verticalement afin de modéliser respectivement, les ligatures horizontales et ses allongements ainsi que les chevauchements des caractères. L'extension de ces modèles à la reconnaissance de noms de villes tunisiennes, s'appuie sur des règles syntaxiques. Mots clés : ligatures horizontales et verticales, chaînes de caractères, durée, PHMM. Plusieurs solutions ont été proposées pour la reconnaissance de l'écriture arabe, nous trouvons dans [2] une liste de travaux déjà réalisés. Comme pour toute écriture cursive, des méthodes globales et analytiques ont été testées. Cependant, à cause de certaines caractéristiques de l'écriture arabe (Tableau 1), la solution n'est pas triviale. En effet, l'étude morphologique de l'arabe montre qu'il est difficile d'opérer la reconnaissance au niveau du caractère et ceci pour les raisons suivantes : Ÿ Le caractère arabe peut avoir jusqu'à 4 formes différentes selon sa position dans le mot. Ÿ Plusieurs groupes de caractères possèdent le même corps mais un nombre et/ou un emplacement de points diacritiques différents. Ces points se situent au dessus ou au dessous de la ligne de base à des endroits différents dans l'image en fonction du caractère et dans certains cas de sa position dans le mot. Ces points sont sensibles au bruit, ils peuvent être accolés au corps ou souvent confondus avec du bruit. Ÿ Certains caractères arabes incluent une boucle qui peut avoir différentes formes. La boucle est très souvent bouchée ou ouverte. Toutes ces raisons nous ont conduit à s'orienter vers la chaîne de caractères, qu'il est convenu d'appeler PAW-Piece of Arabic Word-; le PAW peut correspondre au mot ou à la partie du mot. Les PAWs présentent une structure facile à isoler (séparation en composantes connexes, extraction de contours...), ils peuvent occuper différentes positions dans le mot sans pour autant changer de formes. De plus, ils caractérisent toute la morphologie de l'écriture arabe: on y trouve des caractères isolés et des caractères ligaturés. Cependant, les PAWs contiennent des ligatures horizontales et verticales bien spécifiques : Ÿ Les ligatures horizontales sont variables, elles se traduisent par un allongement de la ligature de base ce qui complique le processus de segmentation. Ÿ Les ligatures verticales, bien qu'elles soient rares, posent d'importants problèmes lorsqu'elles existent. Le chevauchement vertical des caractères modifie souvent la morphologie de certains d'entre eux et la ligne de base n'est plus horizontale. Au vu de ces problèmes, nous avons opté pour une modélisation globale au niveau du PAW. Cependant, ce choix est pénalisé par le nombre important de PAWs dans le cadre d'un vocabulaire libre : il serait inconcevable d'envisager un

In recent years, the established link between the various human communication production domains has become more widely utilised in the field of speech processing. In this work, a state of the art Semi Adaptive Appearance Model (SAAM) approach developed by the authors is used for automatic lip tracking, and an adapted version of our vowel based speech segmentation system is employed to automatically segment speech. Canonical Correlation Analysis (CCA) on segmented and non segmented data in a range of noisy speech environments finds that segmented speech has a significantly better audiovisual correlation, demonstrating the feasibility of our techniques for further development as part of a proposed audiovisual speech enhancement system.

Connected Word Recognition (CWR) is used in many applications such as voice-dialing telephone, automatic data entry, automated banking systems and, etc. This paper presents a novel architecture for CWR based on synergic Hidden Markov Model (HMM) and Dynamic Time Warping (DTW). At first, the proposed architecture eliminates obvious silent times from inputted speech utterance by preprocessing operations. Then, in order to determine boundaries of the existing words in the compressed utterance, a set of candidates for boundary of each word is computed by using the existing capability of the HMM model. Finally, recognition operation is performed by using the synergic between HMM and DTW methods. The architecture has been compared with TLDP method from recognition accuracy and time complexity viewpoints. The evaluation results show that the proposed method significantly improves recognition accuracy and recognition time in comparison with the TLDP method.

To automate assessments of beginning readers, especially those still learning English, we have investigated the types of knowledge sources that teachers use and have tried to incorporate them into an automated system. We describe a set of speech recognition and verification experiments and compare teacher scores with automatic scores in order to decide when a novel pronunciation is best viewed as a reading error or as dialect variation. Since no one classroom teacher is expected to be familiar with as many dialect systems as might occur in an urban classroom, making progress in automated assessments in this area can improve the consistency and fairness of reading assessment. We found that automatic methods performed best when the acoustic models were trained on both native and non-native speech, and argue that this training condition is necessary for automatic reading assessment since a child's reading ability is not directly observable in one utterance. We also found assessment of emerging reading skills in young children to be an area ripe for more research!