Paper accepted at EMNLP 2026 Industry Track

🎉 Our paper “Perceptually Better, Semantically Worse: Measuring Speech Enhancement Impact on LLM-Based Voice Systems” (with Pejman Mowlaee) was accepted at the EMNLP 2026 Industry Track.

Speech enhancement improves perceptual metrics such as PESQ and STOI, but can change what a downstream LLM understands. We introduce the Output Divergence Rate (ODR) and show, for example, that MetricGAN+ more than doubles ODR (0.318 vs. 0.135) despite improving PESQ.

📦 Code and evaluation pipeline: se-llm-odr-benchmark