Expert evaluation of English speech works well on gaming PC
Expert L2 English Speech Evaluation on a Gaming PC: Accuracy, Fluency, and Prosody with XLS-R, XEUS, and PhoneticXeus
Computation and LanguageSound
Summary
Evaluating how well someone speaks a language, like English, often needs experts and special equipment. This paper shows a way to assess pronunciation, fluency, and rhythm of English spoken by learners using a regular gaming computer. The system uses existing speech models combined with trained regression methods and matches or beats a well-known cloud service without needing extra fine-tuning on the test data. This means people can get expert-level feedback on speech using just their personal computers.
What this means in practice
- •For language learning app developers: Integrate expert-level evaluation of English pronunciation, fluency, and prosody into offline language learning apps to provide better learner feedback.$Commercial implications: Enables selling enhanced, on-device pronunciation evaluation features that avoid reliance on cloud services.
- •For speech therapy clinics: Use consumer-grade PCs to deliver objective, expert-informed assessments of English speech skills without expensive or cloud-dependent tools.
Authors
Richard Rose
Abstract
oollim is a locally executable system for evaluating L2 English pronunciation accuracy, fluency, and prosody on a consumer computer. The system combines pretrained XLS-R, XEUS, and PhoneticXeus representations with reference-text features and ridge regressions trained on SpeechOcean762, the Korean Learners' L2 Speech Corpus (English), J-AESOP, and their combinations. On 305 independently held-out recordings assessed by three expert raters, several ridge configurations substantially outperform archived Azure API scores obtained in January 2026 in both Spearman correlation and quadratic weighted kappa across all three constructs. These results support expert speech evaluation on consumer hardware using pretrained representations and construct-specific regressions, without neural-model fine-tuning or calibration on the private evaluation data.