New study: Generative speech enhancement risks hallucination vs discriminative

A comprehensive study on arXiv (2606.02913) by researchers Shrishti Saha Shetu, Emanuël A. P. Habets, and Andreas Brendel pits generative deep learning models (e.g., diffusion-based) against discriminative models (e.g., spectral masking) for speech enhancement. The evaluation covers noise reduction under both high and low signal-to-noise ratio conditions, as well as matched and mismatched training scenarios. The authors also examine the impact of training data volume, model convergence speed, and complexity-performance trade-offs. Notably, they introduce a novel metric: hallucination characteristics, measured via word error rate and phoneme similarity, to quantify when generative models introduce false or distorted speech content.

Results show that generative approaches excel at perceptual quality—producing cleaner, more natural-sounding audio—but at the cost of higher hallucination rates, especially in low-SNR or mismatched conditions. Discriminative models, while less perceptually pleasing, are more robust and less prone to inventing speech. The paper concludes that the perceptual gains of generative methods may not always justify the added computational complexity and hallucination risk, offering practical guidance for engineers deploying speech enhancement in real-world applications like hearing aids, voice assistants, and teleconferencing.