Features For Speech Recognition Research Articles

Transformer-Based Regression Models for Assessing Reading Passage Complexity: A Deep Learning Approach in Natural Language Processing

Natural Language Processing (NLP) is a vital area in deep learning, widely applied in tasks like text classification, virtual assistants, speech recognition, and autocorrect features in digital devices. It allows machines to understand and generate human language, enhancing user interactions with software. This paper presents a deep learning model using the Transformer architecture for a regression task to predict the complexity of reading passages based on text excerpts. By leveraging the Transformer’s capability to identify complex patterns in text, the model achieves a relative error rate of about 10%. The paper also examines how different architectural choices influence model performance, focusing on one-hot encoding and embeddings. While one-hot encoding provides a simple text representation, embeddings offer a richer, more nuanced understanding of word relationships. The findings highlight the significance of model design and data representation in optimizing NLP tasks, providing insights for future advancements in the field.

Nepali Virtual AI Assistant

With the development of Artificial Intelligence (AI), much of our regular redundant work that does not require critical thinking is replaced by machines. The scope of the Internet of Things (IoT) is also increasing due to AI. IoT acts as a suitable platform for the application of AI in real life. In this paper, we present a Virtual AI Assistant that combines AI with IoT. Our system uses an ESP8266 (Node MCU) chip as the core microcontroller for the home automation application. It is an IoT-compatible microcontroller with a cloud connectivity feature that is accomplished with the help of WiFi (Wireless fidelity). This allows home appliances such as lights, fans, rice cookers, etc. to be controlled by desktop or mobile applications from any part of the world. Our system uses speech recognition as its primary source for controlling the application. The speech recognition feature can detect and distinguish people’s voices which is a handy feature in terms of security and privacy. Other than IoT, software features such as note-making, application initiators, web navigation, and automatic typing through speech recognition are also implemented in this system. Our work also makes use of the Nepali language for IoT applications for switching home appliances on and off. The AI model was trained from scratch manually for the Nepali language as proper audio datasets for the Nepali language are not available. Our work in the development of Nepali Virtual AI Assistant is to integrate hardware and software with AI to create user-friendly automation and ease in people’s daily lives.

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.

Seamless M4T for librarians to communicate and provide multilingual collection services

Purpose The purpose of this paper is to explore the use of artificial intelligence (AI) in libraries, especially university libraries, which are faced with users from various countries who have different languages and cultures. Seamless M4T, which is being developed, has great potential for helping university librarians maximize library services by providing ease of communication. Design/methodology/approach Analyzing the possibility of developing Seamless M4T using natural language processing techniques and how to train language models to be smarter AI tools and can be used to break down language barriers between librarians and users. Findings The implementation of AI-based application Seamless M4T can help university librarians provide maximum service to users who are hampered by language and culture with advanced communication skills. Seamless M4T has an automatic speech recognition feature for dozens of languages, so it can translate speech-to-text, text-to-speech or both text and speech. To convert written words into verbal forms, this AI can also translate and transcribe text and speech in real-time without significant delays. Originality/value This paper emphasizes the use of AI in university libraries to improve services, especially in communication due to language differences between librarians and users. Advantages in using AI in libraries can support the collaboration and scholarly communication process.

Self-Supervised ASR Models and Features for Dysarthric and Elderly Speech Recognition

Self-Supervised ASR Models and Features for Dysarthric and Elderly Speech Recognition

Innovative Approach to Teaching EFL through a Game-Based Speaking Mobile Application

Despite the growing body of research on mobile game-based applications, there is a notable lack of studies on teacher-developed game-based mobile applications. This study employed a research and development framework to explore the creation and implementation of a teacher-developed mobile game-based application aimed at improving speaking abilities. The study was conducted in three phases: a needs analysis with 287 participants (123 high school and 164 college students), the design and implementation of the application incorporating 21st-century skills, task-based language teaching, mobile-assisted language learning, automatic speech recognition, and game-like elements, and the final implementation with 16 participants (8 high school and 8 college students). Data were collected through surveys (for needs analysis and the Unified Theory of Acceptance and Use of Technology (UTAUT)), preand post-speaking tests, and interviews. Analysis methods included descriptive statistics, the Wilcoxon signed-rank test, the Mann-Whitney U-test, and thematic analysis. The findings revealed three significant insights: First, both high school and college students believed in the application’s efficacy in enhancing English language proficiency. Second, the SPEAK APP significantly improved speaking abilities, particularly in fluency and pronunciation, for both groups. Third, in-depth interviews highlighted the app’s key features and benefits, noting increased student self-confidence and suggesting improvements. Critical features identified included elements that promote critical thinking, enable experience sharing, facilitate English communication, and encourage interactive peer play. While the application was praised for its user-friendliness, concerns regarding the response time and accuracy of the speech recognition feature indicated areas for future refinement to enhance language learning outcomes.

A novel attention model across heterogeneous features for stuttering event detection

Stuttering is a prevalent speech disorder affecting millions worldwide. To provide an automatic and objective stuttering assessment tool, Stuttering Event Detection (SED) is under extensive investigation for advanced speech research and applications. Despite significant progress achieved by various machine learning and deep learning models, SED directly from speech signal is still challenging due to stuttering speech’s heterogeneous and overlapped nature. This paper presents a novel SED approach using multi-feature fusion and attention mechanisms. The model utilises multiple acoustic features extracted based on different pitch, time-domain, frequency domain, and automatic speech recognition feature to detect stuttering core behaviours more accurately and reliably. In addition, we exploit both spatial and temporal attention mechanisms as well as Bidirectional Long Short-Term Memory (BI-LSTM) modules to learn better representations to improve the SED performance. The experimental evaluation and analysis convincingly demonstrate that our proposed model surpasses the state-of-the-art models on two popular stuttering datasets, with 4% and 3% overall F1 scores, respectively. The superior results indicate the consistency of our proposed method, supported by both multi-feature and attention mechanisms in different stuttering events datasets.

Students’ perception on the use of the Duolingo application as a medium for developing university-level English language skills

Duolingo has been a popular application for language learning nowadays. Using survey design, this study aims to investigate students' perceptions of the Duolingo app for improving English skills at a university in Indonesia. Over half the students surveyed actively use Duolingo. The findings reveals that 51 of the users poses positive attitudes towards the app. Students found that the app was most useful for enhancing listening and speaking skills, including pronunciation and fluency, due to Duolingo's interactive listening exercises and speech recognition features. Gamification elements like points and badges also made learning more enjoyable and maintained student motivation. These audio-focused, gamified, interactive features are what enabled Duolingo to build university students' oral/aural English skills and communicative competence. Additionally, students appreciated Duolingo’s ability to facilitate mobile learning and simple, self-paced studying.. In summary, findings indicate Duolingo's audio-focused, gamified, interactive features have potential for building university students' oral/aural English skills and communicative competence.

5g web service based on AI speech recognition platform in blended teaching application of English in higher education

Abstract This paper first investigates English teaching using an AI speech recognition platform. In the feature value analysis module, using the hidden Markov model is beneficial to improve the accuracy of MFCC features for parameter speech recognition. Therefore, the hidden Markov model is proposed for English audio recognition. Secondly, according to the recognition platform pronunciation error detection, the speech pronunciation scoring algorithm is proposed, and an example analysis of English blended teaching based on AI speech recognition is conducted. The results show that the blended teaching mode is 0.33 points higher than the traditional teaching mode in the entrance examination, 2.96 points higher in the total score of Grade 4, 2.10 points higher in listening, 0.57 points higher in reading, and 0.31 points higher in writing, and the effect of blended teaching is positively correlated with the students’ entrance English level. This study proposes a blended English teaching strategy that utilizes a speech recognition platform to achieve innovation and diversification of English teaching efforts.

The Power-Normalized Cepstral Coefficient (PNCC) for convolutional neural networks-based robust speech command recognition

While implementations of speech recognition grow rapidly in recent years and are slowly being integrated into our daily devices, the problem of noise robustness is still a challenging task, even with the recent advancement of deep learning technologies for speech recognition. The presence of noise may cause a mismatch between training, which is performed in clean conditions, and noisy testing conditions. This paper proposes a method to extract features for speech recognition by employing features derived under the power law scale, i.e., the Power-Normalized Cepstral Coefficient (PNCC). The power-law can provide better compression in low-energy regions so that it is not sensitive when the speech signal is distorted by noise. The features are implemented on speech recognition based on Convolutional Neural Networks (CNNs). The experiments were carried out by TensorFlow’s Speech Command Dataset mixed with various signal-to-noise ratio to evaluate the method. The experimental findings indicate that the accuracy ranges from 81% to 86%.

Deep convolutional neural networks-based features for Indonesian large vocabulary speech recognition

There are great interests in developing speech recognition using deep learning technologies due to their capability to model the complexity of pronunciations, syntax, and language rules of speech data better than the traditional hidden Markov model (HMM) do. But, the availability of large amount of data is necessary for deep learning-based speech recognition to be effective. While this is not a problem for mainstream languages such as English or Chinese, this is not the case for non-mainstream languages such as Indonesian. To overcome this limitation, we present deep features based on convolutional neural networks (CNN) for Indonesian large vocabulary continuous speech recognition in this paper. The CNN is trained discriminatively which is different from usual deep learning implementations where the networks are trained generatively. Our evaluations show that the proposed method on Indonesian speech data achieves 7.26% and 9.01% error reduction rates over the state-of-the-art deep belief networks-deep neural networks (DBN-DNN) for large vocabulary continuous speech recognition (LVCSR), with Mel frequency cepstral coefficients (MFCC) and filterbank (FBANK) used as features, respectively. An error reduction rate of 6.13% is achieved compared to CNN-DNN with generative training.

Combined spectral and speech features for pig speech recognition.

The sound of the pig is one of its important signs, which can reflect various states such as hunger, pain or emotional state, and directly indicates the growth and health status of the pig. Existing speech recognition methods usually start with spectral features. The use of spectrograms to achieve classification of different speech sounds, while working well, may not be the best approach for solving such tasks with single-dimensional feature input. Based on the above assumptions, in order to more accurately grasp the situation of pigs and take timely measures to ensure the health status of pigs, this paper proposes a pig sound classification method based on the dual role of signal spectrum and speech. Spectrograms can visualize information about the characteristics of the sound under different time periods. The audio data are introduced, and the spectrogram features of the model input as well as the audio time-domain features are complemented with each other and passed into a pre-designed parallel network structure. The network model with the best results and the classifier were selected for combination. An accuracy of 93.39% was achieved on the pig speech classification task, while the AUC also reached 0.99163, demonstrating the superiority of the method. This study contributes to the direction of computer vision and acoustics by recognizing the sound of pigs. In addition, a total of 4,000 pig sound datasets in four categories are established in this paper to provide a research basis for later research scholars.

Survey on optimization algorithms in speech processing

Nature is a tremendous provenance of resolving hard and complex problems that exist in the field of computer science because it reveals a very diverse, robust, dynamic, and interesting phenomenon. It constantly finds the optimal solution to resolve its problem which accomplishes exact equity among its element. Nature-inspired algorithms are the heuristic high-level procedure that interprets nature to solve the optimization problem which popularized in the new era of computation. The main objective of this paper is to evaluate the modern technology and enhancement in the nature-inspired algorithm, especially in the application of speech processing, speech recognition, and speech feature selection problems. This paper presents broad collections of global optimization algorithms which have been successfully applied to generate recognition systems that are integrated with metaheuristic algorithms.

Enhanced Deep Hierarchal GRU & BILSTM using Data Augmentation and Spatial Features for Tamil Emotional Speech Recognition

Enhanced Deep Hierarchal GRU & BILSTM using Data Augmentation and Spatial Features for Tamil Emotional Speech Recognition

English Speech Feature Recognition-Based Fuzzy Algorithm and Artificial Intelligent

It is necessary to study the application of digital technology in English speech feature recognition. This paper combines the actual needs of English speech feature recognition to improve the digital algorithm. Moreover, this paper combines fuzzy algorithm to analyze English speech features, analyzes the shortcomings of traditional algorithms, proposes the fuzzy digitized English speech recognition algorithm, and builds an English speech feature recognition model on this basis. In addition, this paper conducts time-frequency analysis on chaotic signals and speech signals, eliminates noise in English speech features, improves the recognition effect of English speech features, and builds an English speech feature recognition system based on digital means. Finally, this paper conducts grouping experiments by inputting students’ English pronunciation forms and counts the results of the experiments to test the performance of the system. The research results show that the method proposed in this paper has a certain effect.

A Recognition Method Based on Speech Feature Parameters-English Teaching Practice

In order to improve the effect of English teaching practice, this paper constructs an intelligent English phonetic teaching system combined with the method of phonetic feature parameter recognition. Moreover, this paper simulates the self-mixing interference signal containing noise by establishing a simulation, analyzes the size of the noise and its various possibilities, and selects the EEMD method as the English speech denoising algorithm. In addition, with the support of an intelligent denoising algorithm, this paper implements an English intelligent teaching system based on the recognition algorithm of English speech feature parameters. Finally, this paper evaluates the teaching effect of the intelligent English speech feature recognition algorithm proposed in this paper and the intelligent teaching system of this paper by means of simulation teaching. The research shows that the English teaching system based on the intelligent speech feature recognition algorithm proposed in this paper has a good effect.

Speaker Adaptation Using Spectro-Temporal Deep Features for Dysarthric and Elderly Speech Recognition

Despite the rapid progress of automatic speech recognition (ASR) technologies targeting normal speech in recent decades, accurate recognition of dysarthric and elderly speech remains highly challenging tasks to date. Sources of heterogeneity commonly found in normal speech including accent or gender, when further compounded with the variability over age and speech pathology severity level, create large diversity among speakers. To this end, speaker adaptation techniques play a key role in personalization of ASR systems for such users. Motivated by the spectro-temporal level differences between dysarthric, elderly and normal speech that systematically manifest in articulatory imprecision, decreased volume and clarity, slower speaking rates and increased dysfluencies, novel spectro-temporal subspace basis deep embedding features derived using SVD speech spectrum decomposition are proposed in this paper to facilitate auxiliary feature based speaker adaptation of state-of-the-art hybrid DNN/TDNN and end-to-end Conformer speech recognition systems. Experiments were conducted on four tasks: the English UASpeech and TORGO dysarthric speech corpora; the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets. The proposed spectro-temporal deep feature adapted systems outperformed baseline i-Vector and x-Vector adaptation by up to 2.63% absolute (8.63% relative) reduction in word error rate (WER). Consistent performance improvements were retained after model based speaker adaptation using learning hidden unit contributions (LHUC) was further applied. The best speaker adapted system using the proposed spectral basis embedding features produced the lowest published WER of 25.05% on the UASpeech test set of 16 dysarthric speakers.

A Pilot Study on the Performance of Time-Domain Features in Speech Recognition based on high-density sEMG.

Features extracted from the surface electromyography (sEMG) signals during the speaking tasks play an essential role in sEMG based speech recognition. However, currently there are no general rules on the optimal choice of sEMG features to achieve satisfactory performance. In this study, a total of 120 electrodes were placed on the face and neck muscles to record the high-density (HD) sEMG signals when subjects spoke ten digits in English. Then ten different time-domain features were computed from the HD sEMG signals and the classification performance of the speech recognition was thoroughly compared. The contribution of each feature was examined by using three performance metrics, which include classification accuracy, sensitivity, and F1-Score. The results showed that, among all the ten different features, the features of WFL, MAV, RMS, and LOGD were considered to be superior because they achieved higher classification accuracies with high sensitivities and higher F1-Scores across subjects/trials in the sEMG-based digit recognition tasks. The findings of this study might be of great value to choose proper signal features that are fed into the classifier in sEMG-based speech recognition.

Speech Recognition using Convolution Deep Neural Networks

The use of a speech recognition model has become extremely important. Speech control has become an important type; Our project worked on designing a word-tracking model by applying speech recognition features with deep convolutional neuro-learning. Six control words are used (start, stop, forward, backward, right, left). Words from people of different ages. Two equal parts, men and women, contribute to our speech dataset which is used to train and test proposed deep neural networks. Collect data in different places in the street, park, laboratory and market. Words ranged in length from 1 to 1.30 seconds for thirty people. Convolutional Neural Network (CNN) is applied as advanced deep neural networks to classify each word from our pooled data set as a multi-class classification task. The proposed deep neural network returned 97.06% as word classification accuracy with a completely unknown speech sample. CNN is used to train and test our data. Our work has been distinguished from many other papers that often use ready-made and fairly consistent data of the isolated word type. While our data are collected in different noisy environments under different conditions and from two types of speech, isolated word and continuous word.

Klasifikasi Dialek Pengujar Bahasa Inggris Menggunakan Random Forest

Speech recognition is one of the important research fields which is currently widely used for various applications. However, speech recognition performance is affected by the dialect of the speaker. Therefore, dialect recognition is often used as an additional feature in speech recognition. The process of recognizing dialects is not easy. Currently, Machine Learning technology is widely applied in dialect recognition. One of the challenges in the introduction of machine learning-based dialects is the imbalance of classes and overlaps in a wide variety of classification techniques. This study applies Random Forest-based oversampling technology for dialect recognition. For hyper-parameter optimization of the random forest algorithm, we apply the Grid Search method. Experiments on Speech Accent Archive data using the MFCC feature resulted in an accuracy of 0.91 and AUC of 0.95