Cascaded-Gaussian Mixture Regression (C-GMR)

What is C-GMR?

Cascaded Gaussian Mixture Regression or C-GMR is a general framework for adapting a GMR (Gaussian Mixture Regression) trained on a large dataset of input-output joint observations, using a limited set of input-only observations. It was originaly developed in the context of speech processing for adapting an acoustic-articulatory inversion GMM trained on a reference speaker, to a new speaker, given a small amount of audio-only observations. In particular, C-GMR framework includes the “Integrated C-GMR” model (IC-GMR) and the Joint-GMR which combine 2 consecutive GMR in a single probabilistic model. For these models, we derived both the exact EM-based training algorithm and inference equation.

How-to-cite

  • Hueber, T., Girin, L., Alameda-Pineda, X., Bailly, G. (2015), “Speaker-Adaptive Acoustic-Articulatory Inversion using Cascaded Gaussian Mixture Regression”, in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 12, pp. 2246-2259
  • Girin, L., Hueber, T., Alameda-Pineda, X., (2017) Extending the Cascaded Gaussian Mixture Regression Framework for Cross-Speaker Acoustic-Articulatory Mapping, in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 25, no. 3, pp. 662-673

Source code

The Matlab source code for training and using the Integrated C-GMR and the Joint-GMR can be downloaded from the C-GMR git repository https://github.com/thueber/cgmr.git

Contributors

  • Dr. Thomas Hueber, CNRS researcher, GIPSA-lab (Grenoble, France)
  • Dr. Laurent Girin, Professor at Grenoble-INP, GIPSA-lab/INRIA (Grenoble, France)
  • Dr. Xavier Alameda-Pineda, INRIA researcher (Montbonnot, France)
  • Dr. Gérard Bailly, CNRS researcher, GIPSA-lab (Grenoble, France)

Contact