Digital document archives are increasingly derived from various different media sources. At present such archives are stored and searched independently. The Information Retrieval from Mixed-Media Collections (IRMMC) project is investigating retrieval from combined document collections composed of items originating from differing media forms. Experimentalin vestigation of a “mixed-media” retrieval task based on the existing TREC Spoken Document Retrieval task combining Text, Spoken and Scanned Image is described. Results show that nontext media perform well within the mixed-media collection. Also while pseudo relevance feedback is extremely effective for spoken documents, its behaviour for document image retrievalis more complex.