A Cocktail-Party Benchmark: Multi-Modal dataset and Comparative Evaluation Results