Automatic Speaker Recognition with Multi-resolution Gaussian Mixture Models ( MR-GMMs )

@inproceedings{DAlmeida2010AutomaticSR,
  title={Automatic Speaker Recognition with Multi-resolution Gaussian Mixture Models ( MR-GMMs )},
  author={Frederico Q. D’Almeida and Francisco Assis de Oliveira Nascimento and Pedro de Azevedo Berger and L{\'u}cio Martins da Silva},
  year={2010}
}
Modern automatic speaker recognition (ASR) systems based on Gaussian Mixture Models (GMMs) have proven quite effective at identifying speakers given certain voice segments (Reynolds, 1992). However, high recognition rates require complex models with at least 16 components (Reynolds and Rose, 1995). If a noise-robust system were developed, then the complexity would likely exceed 80 components (D’Almeida et al., 2008; Ming et al., 2007) and carry a proportional increase in computational costs.