A continuously autoregressive approach for text to motion generation with gaussian mixture-guided latent sampling.