Learning Sequential Latent Variable Models from Multimodal Time Series Data