Intent vs. Surface: Recovering Acoustic Realization from Modern ASR for Pronunciation Training

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Speech Recognition · Depth: Expert, quick

Summary

A study by Seongjin Park investigates whether modern Automatic Speech Recognition (ASR) systems are suitable for pronunciation feedback in language learning, identifying a tendency called "intent bias." This bias causes ASR to recover intended words rather than actual pronunciations, masking mispronunciations. Evaluating eight ASR systems across three architectures on L2-ARCTIC and speechocean762 L2 English corpora, the research found that ASR systems with lower Word Error Rate (WER) inversely correlate with overcorrection, meaning they mask more pronunciation errors. The paper proposes "surface-faithful reranking," an inference-time method using phoneme-level acoustic similarity. This method reduces the false acceptance rate of mispronunciations by 6.0 percentage points on L2-ARCTIC and 5.6 on speechocean762, consistently across diverse learner groups, though substantial overcorrection persists.

Key takeaway

For NLP Engineers developing language learning applications, modern ASR systems, despite their low WER, are inherently suboptimal for pronunciation training due to "intent bias." You should consider implementing surface-faithful reranking to significantly improve mispronunciation detection. Furthermore, advocate for or develop ASR systems with pronunciation-aware objectives to address the remaining overcorrection and enhance feedback accuracy for learners.

Key insights

Modern ASR's intent bias masks mispronunciations, but surface-faithful reranking improves error detection for language learners.

Principles

• ASR systems with lower WER mask more pronunciation errors.
• Pronunciation feedback requires acoustic realization, not just intended words.

Method

Surface-faithful reranking is an inference-time method that uses phoneme-level acoustic similarity to select N-best hypotheses closer to actual learner pronunciation without retraining.

In practice

• Apply surface-faithful reranking to existing ASR systems.
• Focus on phoneme-level acoustic similarity for error detection.

Topics

• ASR
• Pronunciation Training
• Language Learning
• Intent Bias
• Surface-faithful Reranking
• Phoneme-level Acoustic Similarity

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Analyzing Error Propagation in Korean Spoken QA with ASR-LLM Cascades — ASR-stage information loss, especially single-character errors in Korean, is the primary bottleneck for ASR-LLM cascade performance.
• Light-weight Pronunciation Assessment via Discrete Speech Token Surprisal — Lightweight pronunciation assessment uses native speech token surprisal and text-guided alignment, avoiding costly labeled learner data.
• Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India — Real-world Indic ASR requires benchmarks with unscripted speech, diverse languages, and flexible spelling.
• Scaling Human and G2P Supervision for Robust Phonetic Transcription — G2P supervision for phonetic transcription has diminishing returns beyond 20-30 hours of human annotation.
• Transcribing speech is never neutral. It shapes power and bias — Language technologies trained on prestige dialects often misrepresent non-standard and Indigenous communication, leading to serious consequences.
• Subtitle-Aligned Fine-Tuning of Whisper for Swiss German ASR: Benchmark Contamination, Convention Mismatch, and an Honest Baseline at 25.6% WER (13.8% cWER) — Benchmark contamination significantly inflates reported ASR performance, especially for dialectal variations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.