Multimodal Language Analysis is a demanding area of research, since it is\nassociated with two requirements: combining different modalities and capturing\ntemporal information. During the last years, several works have been proposed\nin the area, mostly centered around supervised learning in downstream tasks. In\nthis paper we propose extracting unsupervised Multimodal Language\nrepresentations that are universal and can be applied to different tasks.\nTowards this end, we map the word-level aligned multimodal sequences to 2-D\nmatrices and then use Convolutional Autoencoders to learn embeddings by\ncombining multiple datasets. Extensive experimentation on Sentiment Analysis\n(MOSEI) and Emotion Recognition (IEMOCAP) indicate that the learned\nrepresentations can achieve near-state-of-the-art performance with just the use\nof a Logistic Regression algorithm for downstream classification. It is also\nshown that our method is extremely lightweight and can be easily generalized to\nother tasks and unseen data with small performance drop and almost the same\nnumber of parameters. The proposed multimodal representation models are\nopen-sourced and will help grow the applicability of Multimodal Language.\n