Earticle

본 논문은 잡음에 강인한 음성인식 성능을 획득하기 위해 generalized Gamma 분포기반의 음성향상 기법을 제안한다. 우수한 음성향상을 위해서 제안된 방식에서는 generalized Gamma분포와 spectral gain floor를 이용한 음성추적 기법에 스펙트럼 최소잡음성분에 의한 희귀적인 평균 스펙트럼 값으로부터 유도되는 잡음추정을 결합하여 음질을 향상시켜 음성인식에 적용하였다. Spectral component, spectral amplitude 그리고 log spectral amplitude에 기반하여 제안된 음성향상 기법을 잡음환경에서의 음성인식에 적용하여 그 성능을 측정하였다.

This paper presents a speech enhancement technique based on generalized Gamma distribution in order to obtain robust speech recognition performance. For robust speech enhancement, the noise estimation based on a spectral noise floor controled recursive averaging spectral values is applied to speech estimation under the generalized Gamma distribution and spectral gain floor. The proposed speech enhancement technique is based on spectral component, spectral amplitude, and log spectral amplitude. The performance of three different methods is measured by recognition accuracy of automatic speech recognition (ASR).

본 논문은 DFT기반의 단일마이크 음성향상 방식에 적용된 두 종류의 generalized-Gamma 분포기반의 음성추정 알고리즘을 비교한다. 음성향상 방식으로서는 최소잡음성분에 의한 회귀적인 평균스펙트럼 값으로부터 유도되는 잡음 추정을 각각 인 경우와 인 경우의 Gamma 분포를 이용한 음성추정 기법에 결합하여 음질을 향상시켰다. 각 방식에 의해 향상된 음성신호를 자동차 환경에서의 음성인식에 적용하여 그 성능을 비교하였다.

This paper compares two speech estimators under a generalized Gamma distribution for DFT-based single-microphone speech enhancement methods. For the speech enhancement, the noise estimation based on recursive averaging spectral values by spectral minimum noise is applied to two speech estimators based on the generalized Gamma distribution using or . The performance of two speech enhancement algorithms is measured by recognition accuracy of automatic speech recognition(ASR) in car noisy environment.

Based on local and global constrains, an improved DTW algorithm was proposed in this paper. Firstly, the collected voice data was trained, so the reference template and the test template are constructed by computing the MFCC feature parameters for the training data and the test data, and furthermore, the DTW algorithm was designed according to the local and global constrains. Finally, the proposed DTW algorithm was used to the pattern matching for the reference and test templates. The experimental results show that the proposed algorithm has the remarkable effect in reducing the recognition time and enhancing the recognition accuracy.

本文基于局部约束条件和全局约束条件，提出了一种改进的DTW算法。首先，对采集到的语音数据进行训练，构造由MFCC特征参数构成的参考模板，同时计算测试数据的MFCC特征参数构成测试模板；其次，根据局部约束条件和全局约束条件形成本文的DTW算法；最后，利用改进的DTW算法对测试模板和参考模板进行模式匹配。实验表明，该方法可以明显缩短识别时间，而且具有较高的识别率。

본 논문에서는 지능형 홈네트워크의 음성제어를 위한 가변어휘 연속음성인식시스템을 개발하였다. 또한 자연스런 음성명령에 대한 인식을 위해 핵심어 기반의 자연스런 연속어휘에 대한 대화형 시나리오를 작성하였고, 핵심어기반의 인식 엔진 및 데이터베이스를 구축하여 인식엔진의 성능을 최적화하였다.

In this paper, the vocabulary-independent continuous speech recognition system for speech control of intelligent home-network is presented. This study suggests a conversational scenario of continuous natural vocabulary based upon keywords for recognition on natural speech command, and a way of optimizing the recognition system by constructing a recognition system and database based upon keywords.

This study aims to enhance kiosk accessibility for digitally vulnerable users by designing and implementing a new type of kiosk interface that integrates AI-based motion and speech recognition technologies with gamification elements. Users can explore functions without touch using hand gestures and naturally learn how to operate the system through immersive interactions with mini-games. To evaluate the system’s effectiveness, a Focus Group Interview (FGI) was conducted with older adults, and a heuristic evaluation was carried out with UI/UX experts. The results showed that the proposed interface effectively lowered entry barriers and encouraged repeated use. However, some improvements were identified in visual guidance elements. This study goes beyond conventional UI improvements and serves as an experimental attempt to innovate kiosk user experience through AI technologies and gamified design.

언어학습에서 어휘는 가장 필수적이고 기본이 되는 단위임에도 불구하고 교육현장에서는 학생들에게 어휘를 지도하고 별도의 학습시간을 제공하는 경우는 매우 드물다. 어휘를 습득한다는 것은 소리 내어 말하고 듣는 과정을 통해 이루어진다. 눈으로 내용을 이해하고 암기하는 전통적인 언어 습득 방식은 분명 한계가 있을 수밖에 없다. 본 논문에서는 학습자 특성을 고려한 인지전략과 음성인식을 기반으로 한 Speaking 중심의 학습 방법을 연구하여 초등 영어 어휘 습득을 위한 인지전략 기반의 Speaking Training system을 설계하고 구현하였으며, 초등학교 5학년 두 개 학급을 선정하여 수준 테스트 후 실험 그룹과 비교 그룹으로 각각 편성하여 분석한 결과 학습자의 동기부여와 성취감을 높임으로써 학습자의 소리 영어 중심의 어휘 습득을 강화할 수 있었고, 학력향상 뿐만 아니라 학습참여도, 과제수행 정도, 흥미도 등의 자기주도적 능력까지도 향상시킬 수 있다는 놀라울만한 성과가 있었다. 본 연구를 통해 학생들의 실용적인 영어 말하기 능력을 향상시킬 것으로 기대한다.

In foreign language, vocabulary is the most essential and fundamental elements. Traditional language learning methods that are to understand and to memorize the English contents can only be obvious limitations. In this paper, we proposed the speaking-centered learning methods based on cognitive strategies and speech recognition considering the learner characteristics. We have designed and implemented the cognitive strategy-based speaking training system for acquisition elementary English vocabulary. We were divided into control group and the experimental group and applied to the system to analyze the learning effect. The result of Analysis, the proposed system is increased motivation and achievement of learners. In addition, the proposed system is improved an academic learning participation, Project accomplish, self-interesting and leadership skills. Through this study, we expect that students improve the ability of practical skills in speaking English.

본 논문에서는 대용량 고립단어 언어처리를 위한 통계적 방안에 대한 연구를 수행하였다. 대표적인 대용량 고립단어인 내비게이션 POI(Point of Interest) 단어는 백만 개 이상의 고립단어로 이루어져 있다. 내비게이션에 활용되고 있는 고립단어는 초성 등을 활용한 POI 검색뿐만 아니라 음성인식에 사용된다. 본 논문에서는 하나의 고립 단어로 간주하는 POI를 의미가 있는 단위 단어 세트로 구성된 연속단어로 변환시키는 알고리즘을 제안한다. 먼저 대용량 고립 단어로 이루어진 내비게이션 POI를 분석하여 음성인식 및 POI 검색에 활용이 될 수 있는 단위 단어 세트를 구하였다. 단위 단어 세트를 구하는 알고리즘은 2-3 음절로 이루어진 POI를 초기 단어 세트로 정의한 후 음절이 증가함에 따라 단위 단어 세트를 갱신하는 방식으로 구성되었다. 2-5 음절로 이루어진 653,939개의 POI를 제안된 방식을 사용하여 174,535개의 단위 단어 세트를 구하였으며 이를 이용하여 단위 단어 세트로 이루어진 연속 단어로 기존의 POI를 재 정의하였다. 이를 활용하여 통계적 언어처리 모델에 적용한 결과 복잡도가 485.73으로 나타났다.

In this paper, we make a study on the statistical approach for language processing in a very large vocabulary isolated word recognizer. The representative system is a navigation software in which POI (point of interest) words consist of more than million isolated vocabularies. Those vocabularies have been used for building an inverted-indexed system for searching the first consonant as well as a speech recognizer. We propose an algorithm in which an isolated POI word can be converted into a continuous sentence consisting of a sequence of words in a unit word dictionary. First, a part of POIs is analyzed to make a unit word dictionary. The initial unit word dictionary consists of POIs with two or three syllables and is updated for POIs with more than 4 syllables. We build the unit vocabulary set of 174,535 after analyzing 653,939 POIs having 2-5 syllables. Finally, the perplexity of 485.73 is obtained with the same POIs after statistical language processing for unigram, bigram, and trigram.

Speech emotion recognition (SER) is a critical field within affective computing, aiming to detect and classify emotional states from speech signals, which vary dynamically over time. These signals encode complex relationships between features at multiple time scales, effectively reflecting a speaker’s emotional state. Despite significant progress, SER faces the persistent challenge of labeled data scarcity, a major obstacle given the data-intensive requirements of deep learning models. This scarcity often results in small, imbalanced datasets that hinder model generalization. Various strategies, including feature selection, data augmentation, domain adaptation, and fusion techniques, have been employed to mitigate these issues. However, comprehensive reviews that critically analyze these methods remain limited. In this paper, we provide an extensive review of these data scarcity strategies in SER, assessing their merits and limitations in terms of efficiency and robustness. Special attention is given to how these strategies enhance the performance of both acoustic and multimodal SER systems when operating on limited datasets. Additionally, we highlight the potential of fusion strategies combined with attention mechanisms as promising solutions to improve convergence and reduce model complexity.

In the evolving field of emotion recognition, which intersects psychology, human?computer interaction, and social robotics, there is a growing demand for more advanced and accurate frameworks. The traditional reliance on single-modal approaches has given way to a focus on multimodal emotion recognition, which offers enhanced performance by integrating multiple data sources. This paper introduces TMNet, an innovative multimodal emotion recognition framework that leverages both speech and Electroencephalography (EEG) signals to deliver superior accuracy. This framework utilizes cutting-edge technology, employing a Transformer model to effectively fuse the CNN-BiLSTM and BiGRU architectures, creating a unified multimodal representation for enhanced emotion recognition performance. By utilizing a diverse set of datasets RAVDESS, SAVEE, TESS, and CREMA-D for speech, along with EEG signals captured via the Muse headband. The multimodal model achieves impressive accuracies of 98.89% for speech and EEG signal processing.

For educational and research purposes, we design and evaluate a Korean speech recognition platform to build a decoder. The platform has an object‐oriented architecture so that researchers can modify the platform easily and evaluate the performance of a recognition algorithm of their interests. The platform has the following functionalities: Noise reduction, speech detection, feature extraction, hidden Markov model (HMM)‐based acoustic modeling, cross‐word modeling, ngram language modeling, n‐best search, word graph generation, and Korean‐specific language processing. The platform can handle both lexical search trees for large vocabulary speech recognition, and finite‐state networks for small‐tomedium vocabulary speech recognition. It performs the word‐dependent n‐best search algorithm with a bigram language model in the first forward search stage, then extracts a word lattice, and finally rescores the lattice with a trigram language model in the second backward search stage. In a large vocabulary continuous speech recognition task, we compare the performance of the platform with HTK and Julius.

Background/Objectives: As the fourth industrial age begins, research on speech recognition is actively being carried out. Methods/Statistical analysis: However, since speech recognition is developed based on the standard language, the recognition rate is lower for people using dialects and unusual tones. Because of these problems, it is more difficult to gather data because seniors and dialect users do not attempt to recognize speech. Improvements/Applications: Therefore, to solve this problem, we have added a specific database and configured a speech recognition system that can be used by people who use dialects and unusual tones.

본 사례 연구는 실시간 음성인식을 결합한 설치 작품의 기술, 언어학과 인터랙티브 아트의 융합을 탐구하는 것을 목표로 한다. 따라서 음성인식 기술이 언어라는 축을 통해 ‘Spect’actor’에게 어떻게 전달할 수 있는지에 대한 예술적 관점과 심도 깊은 이해를 도모하고자 한다. 본 연구는 음성인식 기술을 통한 미디어아트 해석 의 가능성을 전반으로 확대하고, 특히 인터랙티브 미디어아트 분야에서 회화적 자동 음성인식을 통한 미디어 아트 해석의 가능성을 넓히고자 합니다.

This case study aims to explore the convergence of technology, linguistics and interactive art of an installation combining real-time speech recognition. It will be followed by fostering an artistic point of view and deeper understanding of how speech recognition technology can convey the ‘spect’actor’ through the axis of language. Therefore, speech recognition can have limitation to media art in space where internet access isn’t furnished nor available. Throughout this research, we will seek to enlarge and inflate large possibilities of interpreting media art throughout spoken language leading to automatic speech recognition, particularly within the field of interactive media art.

Our research aims to enhance the modeling of speech signals for more effective extraction of node features and analysis of relationships between nodes. To achieve this, we model speech signals as cyclic or linear graphs. Our model combines layers of Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to leverage their respective strengths in processing graph data. Specifically, we utilize GCN to aggregate information from neighboring nodes, which helps capture local relationships among nodes. Additionally, we employ GAT mechanisms to assign varying attention weights to different neighboring nodes, facilitating a better capture of complex global relationships between nodes. In our experiments, we validate our approach using the IEMOCAP dataset and demonstrate comparable performance to state-ofthe- art models in emotion recognition tasks. This research outcome provides new insights and methodologies for further exploration in the field of speech signal processing.

Automated speech emotion recognition (SER) by efficient long-term temporal context modeling is a challenging task of the digital audio signal processing domain. However, by default, the recurrent neural network (RNN) is employed to incorporate the temporal dependencies in sequence to investigate the relationships among sequences and features. In this study, we design a parallel convolutional neural network (PCNN) for SER by using a squeeze and excitation network (SEnet) with the self-attention module. Additionally, we adopt the residual learning strategy in both module, SEnet and self-attention, which is further improve the performance of the network. Our proposed SER system utilizes speech spectrogram as input and extracts utterancelevel discrete features by using the PCNN model. We experimentally evaluated our proposed system by standard speech corpus, interactive emotional dyadic motion capture (IEMOCAP). The prediction result reveals the significance and robustness of the proposed PCNN system, which obtained a high recognition rate of 72.01% over state-of-the-art (SOTA) methods.

Voice language, the primary way of human communication, delivers not only verbal information but also emotional information through various characteristics such as voice intonation, height, and surrounding environment. Currently, many studies focus on grasping emotion and speech recognition on voice for human-computer interaction and are developing deep neural network models by extracting various frequency characteristics of speech. Representatives of these speech-based deep learning algorithms include speech recognition, namely speech-to-text or automatic speech recognition, and speech emotion recognition. The development of these two algorithms has been developed for a long time, but multi-output algorithms that process them in parallel at the same time are rare. This paper introduces a multi-output model that recognizes speech and emotion in one voice, thinking that simultaneously understanding language and emotion, which are the most critical information in a human voice, will significantly help human-computer interaction. This model confirmed that there was no significant difference between the training of the language and emotion recognition models separately, with a word error rate of 6.59% in the speech recognition section and an accuracy of 79.67% on average in the emotion recognition section.

Both human listeners and Automatic Speech Recognition (ASR) systems tend to struggle more with recognizing second-language (L2) speech than first-language (L1) speech. This study examined the performance of Whisper (a state-of-the-art ASR system) and L1 English listeners in recognizing L1 and L2 English under a controlled, homogeneous setting (using the same sentences and data collection procedures), enabling a direct comparison across listener and talker types. Speech recordings from 67 L2 English talkers and 25 L1 English talkers embedded in varying levels of background noise were tested, and transcriptions from Whisper and humans were analyzed. Across both L1 and L2 speech, Whisper showed overall higher word recognition accuracy than humans. Notably, it achieved near-perfect performance in quiet or low-noise conditions. Despite this superior performance, Whisper showed greater hallucination rates than humans under loud-noise conditions, with a particularly large gap for L2 speech. Further analysis revealed that Whisper’s hallucination rates remained higher for L2 than L1 speech even after controlling for accuracy, suggesting that these hallucinations are not merely a byproduct of recognition difficulty but reflect a functional difference in how Whisper processes L1 vs. L2 speech. Overall, these findings underscore both the strengths and limitations of Whisper: its robustness in clean listening conditions, but also its hallucination bias against L2 speech.