Speech Classification Research Articles

Analysis of K-NN with the Integration of Bag of Words, TF-IDF, and N-Grams for Hate Speech Classification on Twitter

Social media has emerged as one of the primary communication channels in the modern world, but it has simultaneously become a platform where hate speech can spread easily. This study attempts to evaluate the performance of a hate speech classification model using the K-Nearest Neighbors (K-NN) algorithm along with various feature extraction techniques, specifically Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TF-IDF), and N-Grams. The dataset used in this study consists of 13169 entries, which represent a diverse range of hate speech examples commonly encountered on social media platforms. In this experimental investigation, we assess the efficacy of the model using each feature extraction technique. The findings reveal that the K-NN model exhibits optimal performance when the k parameter is set to 3 (k=3). Under this configuration, the model achieves an accuracy of 86.88%, with a precision of 88.27%, a recall of 86.88%, and an F1-Score of 86.50%. These results show that the integration of TF-IDF feature extraction technique with K-NN algorithm produces superior performance in hate speech classification.

Spectral Similarity Estimation and ML-Based Classification of Speech Using PDMS-Embedded FBG Sensor-Based Contact Microphone

Spectral Similarity Estimation and ML-Based Classification of Speech Using PDMS-Embedded FBG Sensor-Based Contact Microphone

Investigating the role of electroencephalogram based rhythmic bands: applications in imagined speech classification and epileptic seizure severity detection

Investigating the role of electroencephalogram based rhythmic bands: applications in imagined speech classification and epileptic seizure severity detection

Functional and pragmatic parameters of speech acts with the explicit mode of ignorance in scientific dialogue

The article is about identifying functional aspects of 1st-person statements with the explicit mode of ignorance (messages like I do not know… / We don’t know… I don’t know / We don’t know ...) in scientific discussions as a genre of scientific discourse. The relevance of the analysis is predetermined by the necessity to study various parameters of discourse (including both epistemic and pragmatic) as closely connected. The main research method is contextual and pragmalinguistic analysis. The sources of the material are transcripts of modern English and Russian scientific discussions in various fields of knowledge. The article identifies the illocutionary properties of ignorant statements, establishes their pragmatic functions, as well as determines their position in the traditional classification of speech acts. Taking into account the illocutionary orientation of references to ignorance makes it possible to identify representative, interrogative, axiological varieties of ignorance messages as well as their subtypes differentiated on the basis of the degree of their expressiveness, their vector of evaluation, the nature of the representative and interrogative role of speech acts with the mode of ignorance: fully expressive and conditionally expressive, positive and negative, preventive and hesitative, etc. The criterion “pragmatic functions” helps us specify additional functions of references to ignorance in scientific dialogue: to indicate a certain degree of the speaker’s epistemic responsibility for their words, to serve as a means of avoiding the answer, a way to prevent criticism, etc. Moreover, the work identifies factors determining the illocutionary nature of ignorance speech acts: the surface structure of the message, the general context of utterance, the position of the ignorance message in the dialogical unity, sincerity and felicity conditions of speech acts formulated as ignorance messages. References to ignorance are concluded to be quite a significant phenomenon in the structure of scientific dialogue, since they perform a number of important functions in it. The analysis scheme can be used to study statements with other explicit modes (knowledge, opinions, etc.) in various types of discourse.

Формирование логического мышления у старших дошкольников с тяжелыми нарушениями речи на основе информационной теории развития психики Л.М. Веккера

<p style="text-align: justify;"><strong>Objectives. </strong>Contemporary methods of development of logical thinking in preschool children with severe speech disorders (SSD) help to smooth out problems of academic failure at school and social disadaptation. A program for the development of verbal and logical thinking in senior preschool children with SSD was tested. The program was based on the information theory of mental development by L.M. Vekker.</p> <p style="text-align: justify;"><strong>Methods. </strong>The implementation of the experimental program for the development of logical thinking included the creation of psychological and pedagogical conditions: modeling of surrounding relationships and processes; consistent introduction of models of different levels of concretization of the abstraction; creation of problem situations between surrounding objects and phenomena. The study involved children with a diagnosis of SSD (7 boys, 2 girls) aged 6 to 7 years (M = 6.5), attending a preschool educational organization. To identify “deficits” of verbal-logical thinking in children with speech disorders, a group of children without speech disorders (4 boys, 5 girls) of the same age (M=6.5) was used. Two measurements were taken — before and after the implementation of the formative program. To diagnose the parameters of verbal logical thinking, the “Methodology for studying the level of children’s readiness for school” by L.A. Yasyukova was used.</p> <p style="text-align: justify;"><strong>Results</strong>. Comparison of the indicators of the level of development of thinking in children with speech disorders with the results of children without speech disorders at the first stage of the study showed differences (p≤0.01) in the following parameters: speech and visual classifications, abstract thinking. At the second — developmental — stage, a positive shift at a reliable statistical level occurred in the parameters of the methodology: speech analysis-synthesis (p≤0.05), selection of antonyms (p≤0.05), speech analogies (p≤0.01), speech classifications (p≤0.01), visual analysis synthesis (p≤0.01), visual classifications (p≤0.01).</p> <p style="text-align: justify;"><strong>Conclusions</strong>. The results of the study showed that children with speech disorders have learned to use symbolic and speech models to understand logical relationships at the level of visual-effective and visual-figurative thinking. According to the indicator visual analysis-synthesis, children with speech disorders demonstrated results even higher than children without speech disorders. Work is planned to further replenish the children’s vocabulary, to include in the program a set of training in different types of judgments and conclusions.</p>

A robust and automatic method for the recognition of speech category in online learning discourse

A robust and automatic method for the recognition of speech category in online learning discourse

Improving Hate Speech Classification Through Ensemble Learning and Explainable AI Techniques

Improving Hate Speech Classification Through Ensemble Learning and Explainable AI Techniques

Speech Masking of Authorship in Conflict-Prone Texts: Linguistic, Legal and Psychological Aspects

The article introduces a comprehensive interdisciplinary approach to authorship speech masking identification and classification as a research domain of the linguistics of lies. The material for the study was a corpus of conflict-prone texts, which, among other things, contained deliberately falsified information about the author. The paper discloses the issues of conflict-prone texts authorship investigation. It presents the description of the main types of speech masking and linguistic ways of their expression. The legal aspects of the concept are revealed, in particular, the description of crimes is given, the compositions of which may include a speech masking of authorship. Some features of the subjective side of the offense (intent, motive) are studied, thus it enables qualifying the speech behavior of a person as a disguise. The psychological aspect of speech masking in the author’s text is highlighted, trends in psychological influence orientation in the context of various types of its implementation are noted. The comprehensive study resulted in the establishment and description of the types of authorship masking. Based on the study of the text array, the linguistic and psychological characteristics of each of the diagnosed types of disguise were identified in terms of legislative norms and relevant law enforcement practice.

Complaints Speech Acts between Customer and Staff at Western Cafe

A complaint is an expression of dissatisfaction about an act, omission, decision, or service that expects of the feedback given by the user. This study focuses on an interaction between customer and server: Western Café In Canggu strongly conveys the interaction using speech act categories. The objective of this study is to ascertain the many classifications of speech acts employed within the context of interactions between restaurant customers and servers. The method and technique used to capture data for this study were qualitative analysis and note-taking. Furthermore, the data were analyzed by using theory: Searle’s theory of speech act (1969) and Austin (1962) is used to define and classify the categories of speech act used by the server and customer. The findings show from every interaction that guest and server made and expressing of directive and expressive. Directive point is the most common illocutionary act used in conversation that how make easier the hearer to understand what the speaker mean.

Hierarchical Sentiment Analysis Framework for Hate Speech Detection: Implementing Binary and Multiclass Classification Strategy

A significant challenge in automating hate speech detection on social media is distinguishing hate speech from regular and offensive language. These identify an essential category of content that web filters seek to remove. Only automated methods can manage this volume of daily data. To solve this problem, the community of Natural Language Processing is currently investigating different ways of hate speech detection. In addition to those, previous approaches (e.g., Convolutional Neural Networks, multi-channel BERT models, and lexical detection) have always achieved low precision without carefully treating other related tasks like sentiment analysis and emotion classification. They still like to group all messages with specific words in them as hate speech simply because those terms often appear alongside hateful rhetoric. In this research, our paper presented the hate speech text classification system model drawn upon deep learning and machine learning. In this paper, we propose a new multitask model integrated with shared emotional representations to detect hate speech across the English language. The Transformer-based model we used from Hugging Face and sentiment analysis helped us prevent false positives. Conclusion. We conclude that utilizing sentiment analysis and a Transformer-based trained model considerably improves hate speech detection across multiple datasets.

Analysis of FastText with Support Vector Machine for Hate Speech Classification on Twitter Social Media

Analysis of FastText with Support Vector Machine for Hate Speech Classification on Twitter Social Media

When Jack isn't Jacques: Simultaneous opposite language-specific speech perceptual learning in French-English bilinguals.

Humans are remarkably good at understanding spoken language, despite the huge variability of the signal as a function of the talker, the situation, and the environment. This success relies on having access to stable representations based on years of speech input, coupled with the ability to adapt to short-term deviations from these norms, e.g. accented speech or speech altered by ambient noise. In the last two decades, there has been a robust research effort focused on a possible mechanism for adjusting to accented speech. In these studies, listeners typically hear 15 - 20 words in which a speech sound has been altered, creating a short-term deviation from its longer-term representation. After exposure to these items, listeners demonstrate "lexically driven phonetic recalibration"-they alter their categorization of speech sounds, expanding a speech category to take into account the recently heard deviations from their long-term representations. In the current study, we investigate such adjustments by bilingual listeners. French-English bilinguals were first exposed to nonstandard pronunciations of a sound (/s/ or /f/) in one language and tested for recalibration in both languages. Then, the exposure continued with both the original type of mispronunciation in the same language, plus mispronunciations in the other language, in the opposite direction. In a final test, we found simultaneous recalibration in opposite directions for the two languages-listeners shifted their French perception in one direction and their English in the other: Bilinguals can maintain separate adjustments, for the same sounds, when a talker's speech differs across two languages.

TF-IDF combined rank factor Naive Bayesian algorithm for intelligent language classification recommendation systems

With the continuous improvement of smart language systems, a large amount of language text data has emerged. How to efficiently and accurately process these text data has become an important challenge. Therefore, a speech classification recommendation system based on improved Naive Bayesian algorithm is proposed. The system first adopts the traditional Bayesian algorithm to classify language texts. The Term Frequency-Inverse Document Frequency and rank factor are combined to increase the weight of feature languages. Then, the classified language texts are combined with the improved algorithm for language classification recommendation. Finally, performance testing and simulation applications are conducted on the system. From the results, in the Gutenberg corpus, the research algorithm had the highest accuracy and completeness, with 98.5 % and 91.6 %, respectively, and the lowest values were 92.6 % and 89.4 %. The average values were 95.5 % and 91.1 %, with an F1 value of about 92.6 %. In the Brown corpus, the average accuracy, completeness, and F1 value of the designed algorithm were 96.2 %, 91.2 %, and 93.2 %, respectively. When the number of online customers reached 1000, the response time of the designed Chinese system was 1.15 s, the classification recommendation accuracy was 95 %, and the system stability was about 83 % on average. The response time of the English system was 0.64 s, the classification recommendation accuracy was 96 %, and the system stability was about 90 % on average. It shows that the designed method can significantly enhance the operation accuracy of the classification recommendation system.

Exploring the effectiveness of reward-based learning strategies for second-language speech sounds.

Adults struggle to learn non-native speech categories in many experimental settings (Goto, Neuropsychologia, 9(3), 317-323 1971), but learn efficiently in a video game paradigm where non-native speech sounds have functional significance (Lim & Holt, Cognitive Science, 35(7), 1390-1405 2011). Behavioral and neural evidence from this and other paradigms point toward the involvement of reinforcement learning mechanisms in speech category learning (Harmon, Idemaru, & Kapatsinski, Cognition, 189, 76-88 2019; Lim, Fiez, & Holt, Proceedings of the National Academy of Sciences, 116, 201811992 2019). We formalize this hypothesis computationally and implement a deep reinforcement learning network to map between environmental input and actions. Comparing to a supervised model of learning, we show that the reinforcement network closely matches aspects of human behavior in two experiments - learning of synthesized auditory noise tokens and improvement in speech sound discrimination. Both models perform comparably and the similarity in the output of each model leads us to believe that there is little inherent computational benefit to a reward-based learning mechanism. We suggest that the specific neural circuitry engaged by the paradigm and links between striatum and superior temporal areas play a critical role in effective learning.

Abusive and Hate speech Classification in Arabic Text Using Pre-trained Language Models and Data Augmentation

Abusive and Hate speech Classification in Arabic Text Using Pre-trained Language Models and Data Augmentation

Cascaded cross-modal transformer for audio–textual classification

Speech classification tasks often require powerful language understanding models to grasp useful features, which becomes problematic when limited training data is available. To attain superior classification performance, we propose to harness the inherent value of multimodal representations by transcribing speech using automatic speech recognition models and translating the transcripts into different languages via pretrained translation models. We thus obtain an audio–textual (multimodal) representation for each data sample. Subsequently, we combine language-specific Bidirectional Encoder Representations from Transformers with Wav2Vec2.0 audio features via a novel cascaded cross-modal transformer (CCMT). Our model is based on two cascaded transformer blocks. The first one combines text-specific features from distinct languages, while the second one combines acoustic features with multilingual features previously learned by the first transformer block. We employed our system in the Requests Sub-Challenge of the ACM Multimedia 2023 Computational Paralinguistics Challenge. CCMT was declared the winning solution, obtaining an unweighted average recall of 65.41% and 85.87% for complaint and request detection, respectively. Moreover, we applied our framework on the Speech Commands v2 and HVB dialog data sets, surpassing previous studies reporting results on these benchmarks. Our code is freely available for download at: https://github.com/ristea/ccmt.

Perception of second language phonetic contrasts by monolinguals and bidialectals: A comparison of competencies.

This study aims to examine the perception of English vowels by Greek monolingual and bidialectal speakers of English as a second language (L2) and assess the predictions of the Universal Perceptual Model (UPM). Adult Cypriot Greek (CG) bidialectal speakers and Standard Modern Greek (SMG) monolingual speakers participated in classification and discrimination tests. The two groups were matched for various linguistic, sociolinguistic, and cognitive factors. Another group of adult English speakers served as controls. Data analysis has been conducted with the use of Bayesian regression models. The results of the discrimination test were predicted by acoustic similarity only to some extent, whereas perceptual similarity predicted most contrasts, confirming the hypotheses of UPM. A crucial finding was that bidialectals outperformed monolinguals in the discrimination of L2 contrasts. The advantage observed in bidialectals could be attributed to the greater flexibility of their speech categories, stemming from exposure to more diverse linguistic input.

A Study of Discriminatory Speech Classification Based on Improved Smote and SVM-RF

The rapid development of the Internet has facilitated expression, sharing, and interaction on social networks, but some speech may contain harmful discrimination. Therefore, it is crucial to classify such speech. In this paper, we collected discriminatory data from Sina Weibo and propose the improved Synthetic Minority Over-sampling Technique (SMOTE) algorithm based on Latent Dirichlet Allocation (LDA) to improve data quality and balance. And we propose a new integration method integrating Support Vector Machine (SVM) and Random Forest (RF). The experimental results demonstrate that the integrated model exhibits enhanced precision, recall, and F1 score by 6.0%, 5.4%, and 5.7%, respectively, in comparison with SVM alone. Moreover, it exhibits the best performance in comparison with other machine learning methods. Furthermore, the positive impact of improved SMOTE and this integrated method on model classification is also confirmed in ablation experiments.

Speechformer-CTC: Sequential modeling of depression detection with speech temporal classification

Speech-based automatic depression detection systems have been extensively explored over the past few years. Typically, each speaker is assigned a single label (Depressive or Non-depressive), and most approaches formulate depression detection as a speech classification task without explicitly considering the non-uniformly distributed depression pattern within segments, leading to low generalizability and robustness across different scenarios. However, depression corpora do not provide fine-grained labels (at the phoneme or word level) which makes the dynamic depression pattern in speech segments harder to track using conventional frameworks. To address this, we propose a novel framework, Speechformer-CTC, to model non-uniformly distributed depression characteristics within segments using a Connectionist Temporal Classification (CTC) objective function without the necessity of input–output alignment. Two novel CTC-label generation policies, namely the Expectation-One-Hot and the HuBERT policies, are proposed and incorporated in objectives on various granularities. Additionally, experiments using Automatic Speech Recognition (ASR) features are conducted to demonstrate the compatibility of the proposed method with content-based features. Our results show that the performance of depression detection, in terms of Macro F1-score, is improved on both DAIC-WOZ (English) and CONVERGE (Mandarin) datasets. On the DAIC-WOZ dataset, the system with HuBERT ASR features and a CTC objective optimized using HuBERT policy for label generation achieves 83.15% F1-score, which is close to state-of-the-art without the need for phoneme-level transcription or data augmentation. On the CONVERGE dataset, using Whisper features with the HuBERT policy improves the F1-score by 9.82% on CONVERGE1 (in-domain test set) and 18.47% on CONVERGE2 (out-of-domain test set). These findings show that depression detection can benefit from modeling non-uniformly distributed depression patterns and the proposed framework can be potentially used to determine significant depressive regions in speech utterances.

Exploring Spectrogram-Based Audio Classification for Parkinson's Disease: A Study on Speech Classification and Qualitative Reliability Verification.

Patients suffering from Parkinson's disease suffer from voice impairment. In this study, we introduce models to classify normal and Parkinson's patients using their speech. We used an AST (audio spectrogram transformer), a transformer-based speech classification model that has recently outperformed CNN-based models in many fields, and a CNN-based PSLA (pretraining, sampling, labeling, and aggregation), a high-performance model in the existing speech classification field, for the study. This study compares and analyzes the models from both quantitative and qualitative perspectives. First, qualitatively, PSLA outperformed AST by more than 4% in accuracy, and the AUC was also higher, with 94.16% for AST and 97.43% for PSLA. Furthermore, we qualitatively evaluated the ability of the models to capture the acoustic features of Parkinson's through various CAM (class activation map)-based XAI (eXplainable AI) models such as GradCAM and EigenCAM. Based on PSLA, we found that the model focuses well on the muffled frequency band of Parkinson's speech, and the heatmap analysis of false positives and false negatives shows that the speech features are also visually represented when the model actually makes incorrect predictions. The contribution of this paper is that we not only found a suitable model for diagnosing Parkinson's through speech using two different types of models but also validated the predictions of the model in practice.