Search results for: speech recognition performance

Authors: Michal Krecichwost, Zauzanna Miodonska, Pawel Badura

Abstract:

The diagnosis of sigmatism is mostly based on the observation of articulatory organs. It is, however, not always possible to precisely observe the vocal apparatus, in particular in the oral cavity of the patient. Speech processing can allow to objectify the therapy and simplify the verification of its progress. In the described study the methodology for classification of incorrectly pronounced phoneme [s] is proposed. The recordings come from adults. They were registered with the speech recorder at the sampling rate of 44.1 kHz and the resolution of 16 bit. The database of pathological and normative speech has been collected for the study including reference assessments provided by the speech therapy experts. Ten adult subjects were asked to simulate a certain type of stigmatism under the speech therapy expert supervision. In the recordings, the analyzed phone [s] was surrounded by vowels, viz: ASA, ESE, ISI, SPA, USU, YSY. Thirteen MFCC (mel-frequency cepstral coefficients) and RMS (root mean square) values are calculated within each frame being a part of the analyzed phoneme. Additionally, 3 fricative formants along with corresponding amplitudes are determined for the entire segment. In order to aggregate the information within the segment, the average value of each MFCC coefficient is calculated. All features of other types are aggregated by means of their 75th percentile. The proposed method of features aggregation reduces the size of the feature vector used in the classification. Binary SVM (support vector machine) classifier is employed at the phoneme recognition stage. The first group consists of pathological phones, while the other of the normative ones. The proposed feature vector yields classification sensitivity and specificity measures above 90% level in case of individual logo phones. The employment of a fricative formants-based information improves the sole-MFCC classification results average of 5 percentage points. The study shows that the employment of specific parameters for the selected phones improves the efficiency of pathology detection referred to the traditional methods of speech signal parameterization.

Keywords: computer-aided pronunciation evaluation, sibilants, sigmatism diagnosis, speech processing

Procedia PDF Downloads 283

Authors: Esha Banerjee, Atul Kumar Ojha, Girish Nath Jha

Abstract:

This study aims to develop an intonation labeled database for Hindi. Although no single standard for prosody labeling exists in Hindi, researchers in the past have employed perceptual and statistical methods in literature to draw inferences about the behavior of prosody patterns in Hindi. Based on such existing research and largely agreed upon intonational theories in Hindi, this study attempts to develop a manually annotated prosodic corpus of Hindi speech data, which can be used for training speech models for natural-sounding speech in the future. 100 sentences ( 500 words) each for declarative and interrogative types have been labeled using Praat.

Keywords: speech dataset, Hindi, intonation, labeled corpus

Procedia PDF Downloads 197

Authors: Khaled Mohammed

Abstract:

This paper presents novel face recognition and identification approaches for the real-time attendance management problem in large companies/factories and government institutions. The proposed uses the Minimum Ratio (MR) approach for employee identification. Capturing the authentic face variability from a sequence of video frames has been considered for the recognition of faces and resulted in system robustness against the variability of facial features. Experimental results indicated an improvement in the performance of the proposed system compared to the Previous approaches at a rate between 2% to 5%. In addition, it decreased the time two times if compared with the Previous techniques, such as Extreme Learning Machine (ELM) & Multi-Scale Structural Similarity index (MS-SSIM). Finally, it achieved an accuracy of 99%.

Keywords: attendance management system, face detection and recognition, live face recognition, minimum ratio

Procedia PDF Downloads 155

Authors: Agastya Pratap Singh

Abstract:

This review paper examines the emerging applications of generative adversarial networks (GANs) in the fields of image recognition and anomaly detection. With the rapid growth of digital image data, the need for efficient and accurate methodologies to identify and classify images has become increasingly critical. GANs, known for their ability to generate realistic data, have gained significant attention for their potential to enhance traditional image recognition systems and improve anomaly detection performance. The paper systematically analyzes various GAN architectures and their modifications tailored for image recognition tasks, highlighting their strengths and limitations. Additionally, it delves into the effectiveness of GANs in detecting anomalies in diverse datasets, including medical imaging, industrial inspection, and surveillance. The review also discusses the challenges faced in training GANs, such as mode collapse and stability issues, and presents recent advancements aimed at overcoming these obstacles.

Keywords: generative adversarial networks, image recognition, anomaly detection, synthetic data generation, deep learning, computer vision, unsupervised learning, pattern recognition, model evaluation, machine learning applications

Procedia PDF Downloads 25

Authors: Ericson O. Alieto, Aprillete C. Devanadera

Abstract:

The main objective of the study is to determine the dominant speech acts in five commemorative speeches of President Duterte. This study employed Speech Act Theory and Discourse analysis to determine how the speech acts features connote the pragmatic meaning of Duterte’s speeches. Identifying the speech acts is significant in elucidating the underlying message or the pragmatic meaning of the speeches. From the 713 sentences or utterances from the speeches, assertive with 208 occurrences from the corpus or 29% is the dominant speech acts. It was followed by expressive with 177 or 25% occurrences, directive accounts for 152 or 15% occurrences. While commisive accounts for 104 or 15% occurrences and declarative got the lowest percentage of occurrences with 72 or 10% only. These sentences when uttered by Duterte carry a certain power of language to move or influence people. Thus, the present study shows the fundamental message perceived by the listeners. Moreover, the frequent use of assertive and expressive not only explains the pragmatic message of the speeches but also reflects the personality of President Duterte.

Keywords: commemorative speech, discourse analysis, duterte, pragmatics

Procedia PDF Downloads 287

Authors: Vesna Kirandziska, Nevena Ackovska, Ana Madevska Bogdanova

Abstract:

The problem of emotion recognition is a challenging problem. It is still an open problem from the aspect of both intelligent systems and psychology. In this paper, both voice features and facial features are used for building an emotion recognition system. A Support Vector Machine classifiers are built by using raw data from video recordings. In this paper, the results obtained for the emotion recognition are given, and a discussion about the validity and the expressiveness of different emotions is presented. A comparison between the classifiers build from facial data only, voice data only and from the combination of both data is made here. The need for a better combination of the information from facial expression and voice data is argued.

Keywords: emotion recognition, facial recognition, signal processing, machine learning

Procedia PDF Downloads 316

Authors: M. Kiran Reddy, K. Sreenivasa Rao

Abstract:

The conventional Hidden Markov Model (HMM)-based speech synthesis system (HTS) uses only a pulse excitation model, which significantly differs from natural excitation signal. Hence, buzziness can be perceived in the speech generated using HTS. This paper proposes an efficient excitation modeling method that can significantly reduce the buzziness, and improve the quality of HMM-based speech synthesis. The proposed approach models the pitch-synchronous residual frames extracted from the residual excitation signal. Each pitch synchronous residual frame is parameterized using 30 wavelet coefficients. These 30 wavelet coefficients are found to accurately capture the perceptually important information present in the residual waveform. In synthesis phase, the residual frames are reconstructed from the generated wavelet coefficients and are pitch-synchronously overlap-added to generate the excitation signal. The proposed excitation modeling method is integrated into HMM-based speech synthesis system. Evaluation results indicate that the speech synthesized by the proposed excitation model is significantly better than the speech generated using state-of-the-art excitation modeling methods.

Keywords: excitation modeling, hidden Markov models, pitch-synchronous frames, speech synthesis, wavelet coefficients

Procedia PDF Downloads 248

Authors: Dzhavidan Zeinalov, Bugra Sen, Firangiz Aslanova

Abstract:

Most text-to-speech models cannot operate well in low-resource languages and require a great amount of high-quality training data to be considered good enough. Yet, with the improvements made in ASR systems, it is now much easier than ever to collect data for the design of custom text-to-speech models. In this work, our work on using the ASR model to collect data to build a viable text-to-speech system for one of the leading financial institutions of Azerbaijan will be outlined. NVIDIA’s implementation of the Tacotron 2 model was utilized along with the HiFiGAN vocoder. As for the training, the model was first trained with high-quality audio data collected from the Internet, then fine-tuned on the bank’s single speaker call center data. The results were then evaluated by 50 different listeners and got a mean opinion score of 4.17, displaying that our method is indeed viable. With this, we have successfully designed the first text-to-speech model in Azerbaijani and publicly shared 12 hours of audiobook data for everyone to use.

Keywords: Azerbaijani language, HiFiGAN, Tacotron 2, text-to-speech, transfer learning, whisper

Procedia PDF Downloads 44

Authors: Malika Yaici, Soraya Aloui, Sara Semchaoui

Abstract:

This paper deals with human activity recognition from sensor data. It is an active research area, and the main objective is to obtain a high recognition rate. In this work, a recognition system based on expert systems is proposed; the recognition is performed using the objects, object states, and gestures and taking into account the context (the location of the objects and of the person performing the activity, the duration of the elementary actions and the activity). The system recognizes complex activities after decomposing them into simple, easy-to-recognize activities. The proposed method can be applied to any type of activity. The simulation results show the robustness of our system and its speed of decision.

Keywords: human activity recognition, ubiquitous computing, context-awareness, expert system

Procedia PDF Downloads 118

Authors: Ainagul Yermekova, Liudmila Goncharenko, Ali Baghirzade, Sergey Sybachin

Abstract:

In this article, we address the problem of Kazakhstan's transition to the Latin alphabet. The transition process started in 2017 and is scheduled to be completed in 2025. In connection with these events, the problem of recognizing the characters of the new alphabet is raised. Well-known character recognition programs such as ABBYY FineReader, FormReader, MyScript Stylus did not recognize specific Kazakh letters that were used in Cyrillic. The author tries to give an assessment of the well-known method of character recognition that could be in demand as part of the country's transition to the Latin alphabet. Three methods of character recognition: template, structured, and feature-based, are considered through the algorithms of operation. At the end of the article, a general conclusion is made about the possibility of applying a certain method to a particular recognition process: for example, in the process of population census, recognition of typographic text in Latin, or recognition of photos of car numbers, store signs, etc.

Keywords: text detection, template method, recognition algorithm, structured method, feature method

Procedia PDF Downloads 186

Authors: Dev Pratap Singh, Harshika Hasija, Ashwini S.

Abstract:

The project aims to provide a touchless, intuitive interface for human-computer interaction, enabling users to control their computers using hand gestures and voice commands. The system leverages advanced computer vision techniques using the Media Pipe framework and OpenCV to detect and interpret real-time hand gestures, transforming them into mouse actions such as clicking, dragging, and scrolling. Additionally, the integration of a voice assistant powered by the speech recognition library allows for seamless execution of tasks like web searches, location navigation, and gesture control in the system through voice commands.

Keywords: gesture recognition, hand tracking, machine learning, convolutional neural networks, natural language processing, voice assistant

Procedia PDF Downloads 10

Authors: Tadesse Anberbir, Felix Bankole, Tomio Takara, Girma Mamo

Abstract:

In the development of a text-to-speech synthesizer, automatic derivation of correct pronunciation from the grapheme form of a text is a central problem. Particularly deriving phonological features which are not shown in orthography is challenging. In the Amharic language, geminates and epenthetic vowels are very crucial for proper pronunciation but neither is shown in orthography. In this paper, we proposed and integrated a morphological analyzer into an Amharic Text-to-Speech system, mainly to predict geminates and epenthetic vowel positions, and prepared a duration modeling method. Amharic Text-to-Speech system (AmhTTS) is a parametric and rule-based system that adopts a cepstral method and uses a source filter model for speech production and a Log Magnitude Approximation (LMA) filter as the vocal tract filter. The naturalness of the system after employing the duration modeling was evaluated by sentence listening test and we achieved an average Mean Opinion Score (MOS) 3.4 (68%) which is moderate. By modeling the duration of geminates and controlling the locations of epenthetic vowel, we are able to synthesize good quality speech. Our system is mainly suitable to be customized for other Ethiopian languages with limited resources.

Keywords: Amharic, gemination, speech synthesis, morphology, epenthesis

Procedia PDF Downloads 87

Authors: Sadiq Aminu, Ahmed Lamido

Abstract:

This research studies Barack Obama’s second inaugural speech using Halliday’s Systemic Functional Grammar (SFG). SFG is a text grammar which describes how language is used, so that the meaning of the text can be better understood. The primary source of data in this research work is Barack Obama’s second inaugural speech which was obtained from the internet. The analysis of the speech was based on the ideational and textual metafunctions of Systemic Functional Grammar. Specifically, the researcher analyses the Process Types and Participants (ideational) and the Theme/Rheme (textual). It was found that material process (process of doing) was the most frequently used ‘Process type’ and ‘We’ which refers to the people of America was the frequently used ‘Theme’. Application of the SFG theory, therefore, gives a better meaning to Barack Obama’s speech.

Keywords: ideational, metafunction, rheme, textual, theme

Procedia PDF Downloads 159

Authors: Zelalem Fantahun

Abstract:

Part-of-speech tagging is the process of assigning a part-of-speech or other lexical class marker to each word into naturally occurring text. Part-of-speech tagging is the most fundamental and basic task almost in all natural language processing. In natural language processing, the problem of providing large amount of manually annotated data is a knowledge acquisition bottleneck. Since, Amharic is one of under-resourced language, the availability of tagged corpus is the bottleneck problem for natural language processing especially for POS tagging. A promising direction to tackle this problem is to provide a system that does not require manually tagged data. In unsupervised learning, the learner is not provided with classifications. Unsupervised algorithms seek out similarity between pieces of data in order to determine whether they can be characterized as forming a group. This paper explicates the development of unsupervised part-of-speech tagger using K-Means clustering for Amharic language since large amount of data is produced in day-to-day activities. In the development of the tagger, the following procedures are followed. First, the unlabeled data (raw text) is divided into 10 folds and tokenization phase takes place; at this level, the raw text is chunked at sentence level and then into words. The second phase is feature extraction which includes word frequency, syntactic and morphological features of a word. The third phase is clustering. Among different clustering algorithms, K-means is selected and implemented in this study that brings group of similar words together. The fourth phase is mapping, which deals with looking at each cluster carefully and the most common tag is assigned to a group. This study finds out two features that are capable of distinguishing one part-of-speech from others these are morphological feature and positional information and show that it is possible to use unsupervised learning for Amharic POS tagging. In order to increase performance of the unsupervised part-of-speech tagger, there is a need to incorporate other features that are not included in this study, such as semantic related information. Finally, based on experimental result, the performance of the system achieves a maximum of 81% accuracy.

Keywords: POS tagging, Amharic, unsupervised learning, k-means

Procedia PDF Downloads 451

Authors: Hirokatsu Kawashima

Abstract:

One important element of teaching/learning listening is intensive listening such as listening for precise sounds, words, grammatical, and semantic units. Several classroom-based investigations have been conducted to explore the usefulness of auditory recognition of phonological, grammatical and semantic errors in such a context. The current study reports the results of one such investigation, which targeted auditory recognition of phonological, grammatical, and semantic errors at the sentence level. 56 Japanese EFL learners participated in this investigation, in which their recognition performance of phonological, grammatical and semantic errors was measured on a 9-point scale by learners’ self-evaluation from the perspective of 1) two types of similar English sound (vowel and consonant minimal pair words), 2) two types of sentence word order (verb phrase-based and noun phrase-based word orders), and 3) two types of semantic consistency (verb-purpose and verb-place agreements), respectively, and their general listening proficiency was examined using standardized tests. A number of findings have been made about the interactive relationships between the three types of auditory error recognition and general listening proficiency. Analyses based on the OPLS (Orthogonal Projections to Latent Structure) regression model have disclosed, for example, that the three types of auditory error recognition are linked in a non-linear way: the highest explanatory power for general listening proficiency may be attained when quadratic interactions between auditory recognition of errors related to vowel minimal pair words and that of errors related to noun phrase-based word order are embraced (R2=.33, p=.01).

Keywords: auditory error recognition, intensive listening, interaction, investigation

Procedia PDF Downloads 513

Authors: Steffy Maria Joseph, Abdu Rahiman V, Abdul Hameed K. M.

Abstract:

Online handwritten character recognition is a challenging field in Artificial Intelligence. The classification success rate of current techniques decreases when the dataset involves similarity and complexity in stroke styles, number of strokes and stroke characteristics variations. Malayalam is a complex south indian language spoken by about 35 million people especially in Kerala and Lakshadweep islands. In this paper, we consider the significant feature extraction for the similar stroke styles of Malayalam. This extracted feature set are suitable for the recognition of other handwritten south indian languages like Tamil, Telugu and Kannada. A classification scheme based on support vector machines (SVM) is proposed to improve the accuracy in classification and recognition of online malayalam handwritten characters. SVM Classifiers are the best for real world applications. The contribution of various features towards the accuracy in recognition is analysed. Performance for different kernels of SVM are also studied. A graphical user interface has developed for reading and displaying the character. Different writing styles are taken for each of the 44 alphabets. Various features are extracted and used for classification after the preprocessing of input data samples. Highest recognition accuracy of 97% is obtained experimentally at the best feature combination with polynomial kernel in SVM.

Keywords: SVM, matlab, malayalam, South Indian scripts, onlinehandwritten character recognition

Procedia PDF Downloads 574

Authors: K. Prabhu, K. Dinesh, M. Ranjani, M. Suhitha

Abstract:

During the tough times of covid, those people who were hospitalized found it difficult to always convey what they wanted to or needed to the attendee. Sometimes the attendees might also not be there. In that case, the patients can use simple hand gestures to control electrical appliances (like its set it for a zero watts bulb)and three other gestures for voice note intimation. In this AI-based hand recognition project, NodeMCU is used for the control action of the relay, and it is connected to the firebase for storing the value in the cloud and is interfaced with the python code via raspberry pi. For three hand gestures, a voice clip is added for intimation to the attendee. This is done with the help of Google’s text to speech and the inbuilt audio file option in the raspberry pi 4. All the five gestures will be detected when shown with their hands via the webcam, which is placed for gesture detection. The personal computer is used for displaying the gestures and for running the code in the raspberry pi imager.

Keywords: nodeMCU, AI technology, gesture, patient

Procedia PDF Downloads 167

Authors: Helmi Baazaoui, Mounir Zrigui

Abstract:

This study addresses the challenge of hate speech detection in Tunisian Arabic text, a critical issue for online safety and moderation. Leveraging the strengths of the AraBERT model, we fine-tuned and evaluated its performance against the Bi-LSTM model across four distinct datasets: T-HSAB, TNHS, TUNIZI-Dataset, and a newly compiled dataset with diverse labels such as Offensive Language, Racism, and Religious Intolerance. Our experimental results demonstrate that AraBERT significantly outperforms Bi-LSTM in terms of Recall, Precision, F1-Score, and Accuracy across all datasets. The findings underline the robustness of AraBERT in capturing the nuanced features of Tunisian Arabic and its superior capability in classification tasks. This research not only advances the technology for hate speech detection but also provides practical implications for social media moderation and policy-making in Tunisia. Future work will focus on expanding the datasets and exploring more sophisticated architectures to further enhance detection accuracy, thus promoting safer online interactions.

Keywords: hate speech detection, Tunisian Arabic, AraBERT, Bi-LSTM, Gemini annotation tool, social media moderation

Procedia PDF Downloads 11

Authors: Ya-Hui Wang

Abstract:

Lower speech therapy rate of tracheostomized patient was noted in comparison with previous researches. This study is aim to shed light on the referral status of speech therapy in those patients in Taiwan. This study developed an analysis for the size and key characteristics of the population of tracheostomized in-patient in the Taiwan. Method: We analyzed National Healthcare Insurance data (The Collaboration Center of Health Information Application, CCHIA) from Jan 1 2010 to Dec 31 2010. Result: over ages 3, number of tracheostomized in-patient is directly proportional to age. A high service loading was observed in North region in comparison with other regions. Only 4.87% of the tracheostomized in-patients were referred for speech therapy, and 1.9% for swallow examination, 2.5% for communication evaluation.

Keywords: refer, speech therapy, training, rehabilitation

Procedia PDF Downloads 440

Authors: Malika Yaici, Soraya Aloui, Sara Semchaoui

Abstract:

Recognition of human activities from sensor data is an active research area, and the main objective is to obtain a high recognition rate. In this work, we propose a recognition system based on expert systems. The proposed system makes the recognition based on the objects, object states, and gestures, taking into account the context (the location of the objects and of the person performing the activity, the duration of the elementary actions, and the activity). This work focuses on complex activities which are decomposed into simple easy to recognize activities. The proposed method can be applied to any type of activity. The simulation results show the robustness of our system and its speed of decision.

Keywords: human activity recognition, ubiquitous computing, context-awareness, expert system

Procedia PDF Downloads 140

Authors: Gamze Uslu, Sebnem Baydere, Alper K. Demir

Abstract:

In this study, data loss tolerance of Support Vector Machines (SVM) based activity recognition model and multi activity classification performance when data are received over a lossy wireless sensor network is examined. Initially, the classification algorithm we use is evaluated in terms of resilience to random data loss with 3D acceleration sensor data for sitting, lying, walking and standing actions. The results show that the proposed classification method can recognize these activities successfully despite high data loss. Secondly, the effect of differentiated quality of service performance on activity recognition success is measured with activity data acquired from a multi hop wireless sensor network, which introduces high data loss. The effect of number of nodes on the reliability and multi activity classification success is demonstrated in simulation environment. To the best of our knowledge, the effect of data loss in a wireless sensor network on activity detection success rate of an SVM based classification algorithm has not been studied before.

Keywords: activity recognition, support vector machines, acceleration sensor, wireless sensor networks, packet loss

Procedia PDF Downloads 475

Authors: Rabab M. Ramadan, Elaraby A. Elgallad

Abstract:

With extensive application, the performance of unimodal biometrics systems has to face a diversity of problems such as signal and background noise, distortion, and environment differences. Therefore, multimodal biometric systems are proposed to solve the above stated problems. This paper introduces a bimodal biometric recognition system based on the extracted features of the human palm print and iris. Palm print biometric is fairly a new evolving technology that is used to identify people by their palm features. The iris is a strong competitor together with face and fingerprints for presence in multimodal recognition systems. In this research, we introduced an algorithm to the combination of the palm and iris-extracted features using a texture-based descriptor, the Scale Invariant Feature Transform (SIFT). Since the feature sets are non-homogeneous as features of different biometric modalities are used, these features will be concatenated to form a single feature vector. Particle swarm optimization (PSO) is used as a feature selection technique to reduce the dimensionality of the feature. The proposed algorithm will be applied to the Institute of Technology of Delhi (IITD) database and its performance will be compared with various iris recognition algorithms found in the literature.

Keywords: iris recognition, particle swarm optimization, feature extraction, feature selection, palm print, the Scale Invariant Feature Transform (SIFT)

Procedia PDF Downloads 235

Authors: Luis Alvarado, Victor Poblete, Isaac Gonzalez, Yetzabeth Gonzalez

Abstract:

In MIREX 2016 (http://www.music-ir.org/mirex), the deep neural network (DNN)-Deep Chroma Extractor, proposed by Korzeniowski and Wiedmer, reached the highest score in an audio chord recognition task. In the present paper, this tool is assessed under acoustic reverberant environments and distinct source-microphone distances. The evaluation dataset comprises The Beatles and Queen datasets. These datasets are sequentially re-recorded with a single microphone in a real reverberant chamber at four reverberation times (0 -anechoic-, 1, 2, and 3 s, approximately), as well as four source-microphone distances (32, 64, 128, and 256 cm). It is expected that the performance of the trained DNN will dramatically decrease under these acoustic conditions with signals degraded by room reverberation and distance to the source. Recently, the effect of the bio-inspired Locally-Normalized Cepstral Coefficients (LNCC), has been assessed in a text independent speaker verification task using speech signals degraded by additive noise at different signal-to-noise ratios with variations of recording distance, and it has also been assessed under reverberant conditions with variations of recording distance. LNCC showed a performance so high as the state-of-the-art Mel Frequency Cepstral Coefficient filters. Based on these results, this paper proposes a variation of locally-normalized triangular filters called Locally-Normalized Quarter Tone (LNQT) filters. By using the LNQT spectrogram, robustness improvements of the trained Deep Chroma Extractor are expected, compared with classical triangular filters, and thus compensating the music signal degradation improving the accuracy of the chord recognition system.

Keywords: chord recognition, deep neural networks, feature extraction, music information retrieval

Procedia PDF Downloads 232

Authors: Mohammed Arif, Abdessalam Kifouche

Abstract:

Speech is a very convenient way of communication between people and machines. It conveys information about the identity of the talker. Since speaker recognition technology is increasingly securing our everyday lives, the objective of this paper is to develop two automatic text-independent speaker verification systems (TI SV) using low-level spectral features and machine learning methods. (i) The first system is based on a support vector machine (SVM), which was widely used in voice signal processing with the aim of speaker recognition involving verifying the identity of the speaker based on its voice characteristics, and (ii) the second is based on Gaussian Mixture Model (GMM) and Universal Background Model (UBM) to combine different functions from different resources to implement the SVM based.

Keywords: speaker verification, text-independent, support vector machine, Gaussian mixture model, cepstral analysis

Procedia PDF Downloads 58

Authors: Svitov David, Alyamkin Sergey

Abstract:

The usage of convolutional neural networks (CNNs) in conjunction with the margin-based softmax approach demonstrates the state-of-the-art performance for the face recognition problem. Recently, lightweight neural network models trained with the margin-based softmax have been introduced for the face identification task for edge devices. In this paper, we propose a distillation method for lightweight neural network architectures that outperforms other known methods for the face recognition task on LFW, AgeDB-30 and Megaface datasets. The idea of the proposed method is to use class centers from the teacher network for the student network. Then the student network is trained to get the same angles between the class centers and face embeddings predicted by the teacher network.

Keywords: ArcFace, distillation, face recognition, margin-based softmax

Procedia PDF Downloads 146

Authors: Djendi Mohmaed

Abstract:

In this paper, we propose a dual adaptive algorithm, which is based on the combination between the forward blind source separation (FBSS) structure and the affine projection algorithm (APA). This proposed algorithm combines the advantages of the source separation properties of the FBSS structure and the fast convergence characteristics of the APA algorithm. The proposed algorithm needs two noisy observations to provide an enhanced speech signal. This process is done in a blind manner without the need for ant priori information about the source signals. The proposed dual forward blind source separation affine projection algorithm is denoted (DFAPA) and used for the first time in an airplane cockpit context to enhance the communication from- and to- the airplane. Intensive experiments were carried out in this sense to evaluate the performance of the proposed DFAPA algorithm.

Keywords: adaptive algorithm, speech enhancement, system mismatch, SNR

Procedia PDF Downloads 135

Authors: Aqsa Ali, Aleem Mushtaq, Attaullah Memon, Monna

Abstract:

In this paper, we present a low cost design for a smart glove that can perform sign language recognition to assist the speech impaired people. Specifically, we have designed and developed an Assistive Hand Gesture Interpreter that recognizes hand movements relevant to the American Sign Language (ASL) and translates them into text for display on a Thin-Film-Transistor Liquid Crystal Display (TFT LCD) screen as well as synthetic speech. Linear Bayes Classifiers and Multilayer Neural Networks have been used to classify 11 feature vectors obtained from the sensors on the glove into one of the 27 ASL alphabets and a predefined gesture for space. Three types of features are used; bending using six bend sensors, orientation in three dimensions using accelerometers and contacts at vital points using contact sensors. To gauge the performance of the presented design, the training database was prepared using five volunteers. The accuracy of the current version on the prepared dataset was found to be up to 99.3% for target user. The solution combines electronics, e-textile technology, sensor technology, embedded system and machine learning techniques to build a low cost wearable glove that is scrupulous, elegant and portable.

Keywords: American sign language, assistive hand gesture interpreter, human-machine interface, machine learning, sensing glove

Procedia PDF Downloads 301

Authors: Ziyad Motala

Abstract:

It is often proclaimed by leasers in Muslim majority countries that Islamic Law does not permit dissent against a ruler. This paper will evaluate and discuss freedom of speech and dissent as found in concrete prophetic examples during the time of the Prophet Muhammad. It will further look at the examples and practices during the time of the four Noble Caliphs, the immediate successors to the Prophet Muhammad. It will argue that the positivist position of absolute obedience to a ruler is inconsistent with the prophetic tradition. The examples of the Prophet and his immediate four successors (whose lessons Sunni Islam considers to be a source of Islamic Law) demonstrates among the earliest example of freedom of speech and dissent in human history. That tradition frowned upon an inert and uninvolved citizenry. It will conclude with lessons for modern day Muslim majority countries arguing with empirical evidence that freedom of speech, dissent and the right to be governed by consensus versus coercion are fundamental requisites of Islamic law.

Keywords: islamic law, demoracy, freedom of speech, right to dissent

Procedia PDF Downloads 75

Authors: I. Schiller, D. Morsomme, A. Remacle

Abstract:

Children’s ability to process spoken language develops until the late teenage years. At school, where efficient spoken language processing is key to academic achievement, listening conditions are often unfavorable. High background noise and poor teacher’s voice represent typical sources of interference. It can be assumed that these factors particularly affect primary school children, because their language and literacy skills are still low. While it is generally accepted that background noise and impaired voice impede spoken language processing, there is an increasing need for analyzing impacts within specific linguistic areas. Against this background, the aim of the study was to investigate the effect of speech-shaped noise and imitated dysphonic voice on first-grade primary school children’s speech perception and sentence comprehension. Via headphones, 5 to 6-year-old children, recruited within the French-speaking community of Belgium, listened to and performed a minimal-pair discrimination task and a sentence-picture matching task. Stimuli were randomly presented according to four experimental conditions: (1) normal voice / no noise, (2) normal voice / noise, (3) impaired voice / no noise, and (4) impaired voice / noise. The primary outcome measure was task score. How did performance vary with respect to listening condition? Preliminary results will be presented with respect to speech perception and sentence comprehension and carefully interpreted in the light of past findings. This study helps to support our understanding of children’s language processing skills under adverse conditions. Results shall serve as a starting point for probing new measures to optimize children’s learning environment.

Keywords: impaired voice, sentence comprehension, speech perception, speech-shaped noise, spoken language processing

Procedia PDF Downloads 192

Authors: Sevil Igit, Merve Meric, Sarp Erturk

Abstract:

In this paper, it is proposed to improve Daisy descriptor based face recognition using a novel One-Bit Transform (1BT) based pre-registration approach. The 1BT based pre-registration procedure is fast and has low computational complexity. It is shown that the face recognition accuracy is improved with the proposed approach. The proposed approach can facilitate highly accurate face recognition using DAISY descriptor with simple matching and thereby facilitate a low-complexity approach.

Keywords: face recognition, Daisy descriptor, One-Bit Transform, image registration

Procedia PDF Downloads 367