A review of pronunciation assessment for non-native english speech = 비원어민 영어 발음 평가에 관한 조사 연구

국문 초록 (Abstract)

자동 발음 평가(Automatic Pronunciation Assessment, APA)는 제2언어 학습자에게 객관적이고 확장 가능한 피드백을 제공하는 핵심 기술로 자리 잡고 있다.본 연구는 비원어민 영어 학습자를 위한 APA 기술의 최근 발전 동향을 체계적으로 고찰하고, 딥러닝(Deep Learning)과 자기지도학습(Self-Supervised Learning, SSL)의 발전에도 불구하고 여전히 남아 있는 주요 한계를 분석하였다.발음 평가의 정확도는 향상되었으나 모델의 해석 가능성 부족, 평가 기준의 일관성 결여, 언어 간 일반화의 한계가 여전히 실용적 적용을 제약한다. 이에 본 논문은 GOP(Goodness of Pronunciation) 기반부터 계층적(Hierarchical) 및 세분화(Granularity-Decoupled) 구조에 이르는 대표적 모델을 비교·분석하고, 자기지도학습 기반 표현(Self-Supervised Representations), 계층적 모델링(Hierarchical Modeling), 표준화된 말뭉치 주석(Corpus Annotation)을 통합한 피드백 지향형 통합 프레임워크(Feedback-Oriented Unified Framework)를 제안한다. 이러한 접근은 정확성, 해석 가능성, 교육적 활용성을 동시에 향상시킬 수 있는 차세대 발음 평가 시스템의 기반을 제시한다.

다국어 초록 (Multilingual Abstract)

Automatic pronunciation assessment has become a crucial technology for providing objective and scalable feedback to second language learners. This study reviews recent developments in APA for non-native English speakers and highlights key challenges that remain despite advances in deep learning and self-supervised learning. While accuracy has improved, persistent issues in model interpretability, scoring consistency, and cross-linguistic generalization still limit practical application. By comparing representative models ranging from GOP-based approaches to hierarchical and granularity-decoupled architectures, this paper identifies structural limitations and proposes a feedback-oriented framework that integrates self-supervised representations, hierarchical modeling, and standardized corpus annotation to enable more interpretable and pedagogically meaningful pronunciation assessment.

목차 (Table of Contents)

• 1. Introduction 1
• 2. System Framework 5
• 2.1 Workflow of an APA System 5
• 2.2 System Modules 6
• 2.2.1 Feature Extraction 7
• 2.2.2 Scoring Module 9
• 3. Data Resources 11
• 3.1 English L2 Corpora 11
• 3.2 Non-English and Low-Resource Corpora 15
• 3.3 Data Challenges and Trends 16
• 4. Research Avenues 18
• 4.1 Traditional Acoustic Models and GOP Framework 18
• 4.1.1 Principles of GOP 18
• 4.1.2 Limitations of GOP 19
• 4.1.3 Improvements of GOP 20
• 4.2 Deep Feature Representations 24
• 4.2.1 ASR Deep Feature 25
• 4.2.2 Self-Supervised Learning Representations 27
• 4.2.3 Novel Feature Engineering 30
• 4.3 Core Modeling Architectures 32
• 4.3.1 Multi-Granularity Scoring 32
• 4.3.2 Baseline Architecture 32
• 4.3.3 Hierarchical Architectures 34
• 4.3.4 Granularity-Decoupled Architecture 39
• 4.3.5 Multi-Encoder Architecture 41
• 4.4 Related Tasks 43
• 4.4.1 Multi-Aspect and Multi-Granularity Assessment 44
• 4.4.2 Assessing Suprasegmentals 44
• 4.4.3 Joint Task Learning 45
• 4.4.4 Automated Speaking Assessment 47
• 4.4.5 Application Scenario 48
• 4.5 Loss Function Design 49
• 4.5.1 Loss Functions for Ordinal Regression and Contrastive Learning 50
• 4.5.2 Loss Functions for Data Imbalance 50
• 4.6 Human Perception and Assessment Subjectivity 51
• 4.6.1 Assessment Criteria 51
• 4.6.2 Modeling Assessor Bias 52
• 4.6.3 Assessing Intelligibility 53
• 4.6.4 Assessment Stability 54
• 4.7 Technical Challenges 55
• 4.7.1 Alignment Issues 55
• 4.7.2 Data Scarcity and Imbalance 57
• 4.7.3 Acoustic Robustness 59
• 4.7.4 L1-Specific and Non-English Assessment 60
• 4.8 Emerging Opportunities 62
• 4.8.1 Toward General-Purpose Models 62
• 4.8.2 Zero-Shot Assessment and Feedback 63
• 4.8.3 Fine-Tuning Strategies 63
• 5. Conclusion 65
• 6. Future Perspectives 68
• References 70
• ABSTRACT 77