Scholarlink

Cloud-based pronunciation assessment in education introduces material risks to student voice-data privacy and compliance under FERPA, COPPA, and GDPR, motivating an on-premises alternative that eliminates external transfer of recordings [6, 7]. This paper presents a fully containerized English pronunciation assessment system that executes end-to-end locally while delivering actionable feedback for instruction [5]. The pipeline standardizes audio to 16 kHz mono WAV, extracts 40-dimensional MFCCs with CMVN and LDA+MLLT, performs phoneme-level forced alignment, and computes 84-dimensional GOP features composed of 42 LPP and 42 LPR components [5]. A compact GOPT-based Transformer consumes GOP sequences to produce phoneme-, word-, and utterance-level scores across accuracy, completeness, fluency, and prosody [1]. The implementation integrates Kaldi, PyTorch, and FastAPI on a Windows 10 CPU host and is packaged for reproducible on-premises deployment [5]. Validation on representative utterances shows robust discrimination across pronunciation quality conditions, supporting the feasibility of privacy-preserving institutional deployment without sacrificing reliability or operability [1].

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Privacy-Preserving Pronunciation Assessment: Implementation and Validation of a GOPT-Based English Pronunciation Assessment System

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mingchun Liang,
  • Enpin Ren,
  • Cheukhang Lee,
  • Cheongyin Choy,
  • Rouchen Liu

摘要

Cloud-based pronunciation assessment in education introduces material risks to student voice-data privacy and compliance under FERPA, COPPA, and GDPR, motivating an on-premises alternative that eliminates external transfer of recordings [6, 7]. This paper presents a fully containerized English pronunciation assessment system that executes end-to-end locally while delivering actionable feedback for instruction [5]. The pipeline standardizes audio to 16 kHz mono WAV, extracts 40-dimensional MFCCs with CMVN and LDA+MLLT, performs phoneme-level forced alignment, and computes 84-dimensional GOP features composed of 42 LPP and 42 LPR components [5]. A compact GOPT-based Transformer consumes GOP sequences to produce phoneme-, word-, and utterance-level scores across accuracy, completeness, fluency, and prosody [1]. The implementation integrates Kaldi, PyTorch, and FastAPI on a Windows 10 CPU host and is packaged for reproducible on-premises deployment [5]. Validation on representative utterances shows robust discrimination across pronunciation quality conditions, supporting the feasibility of privacy-preserving institutional deployment without sacrificing reliability or operability [1].