Speaker diarization performance can degrade when audio quality is poor or different speakers have similar voices. This work estimates the Diarization Error Rate (DER) directly from audio without requiring ground-truth annotations. The method combines audio-quality features derived from disagreement between two voice activity detection systems and DNSMOS with speaker-discriminability features based on the Silhouette Score and Davies-Bouldin Index. A support vector regression model maps these features to a predicted DER. Experiments on MSDWild and VoxConverse show strong correlations between the estimated and ground-truth DERs, with Pearson correlation coefficients of 0.806 and 0.800, respectively.