| Paper: | SLP-P5.7 | ||
| Session: | Feature-based Robust Speech Recognition | ||
| Time: | Tuesday, May 16, 16:30 - 18:30 | ||
| Presentation: | Poster | ||
| Topic: | Speech and Spoken Language Processing: Feature-based Robust Speech Recognition (e.g., noise, etc) | ||
| Title: | Evaluation of the SPACE denoising Algorithm on Aurora2 | ||
| Authors: | Christophe Cerisara, INRIA-LORIA, France; Khalid Daoudi, IRIT / CNRS, France | ||
| Abstract: | In the area of robust automatic speech recognition, we have introduced recently the SPACE algorithm that denoises speech by mapping two GMMs, which respectively model clean and noisy speech. Each Gaussian of the noisy GMM corresponds to a Gaussian of the clean GMM. In this work, we evaluate SPACE on Aurora2 and identify some weaknesses, which relate to the correspondance between the clean and noisy GMMs. We thus propose a new training procedure for the GMMs that improves this correspondance. We further develop a new algorithm that adapts the noisy GMM to an unknown environment, and preserves its correspondance with the clean GMM. This adapted systems outperforms the multistyle models on the three test sets of Aurora2. | ||
©2018 Conference Management Services, Inc. -||- email: webmaster@icassp2006.org -||- Last updated Friday, August 17, 2012