Experiments on Adaptation Methods to Improve Acoustic Modeling for French Speech Recognition

To improve the performance of Automatic Speech Recognition (ASR) systems, the models must be retrained in order to better adjust to the speaker’s voice characteristics, the environmental and channel conditions or the context of the task. In this project we focus on the mismatch between the acoustic features used to train the model and the vocal characteristics of the front-end user of the system. To overcome this mismatch, speaker adaptation techniques have been used. A significant performance improvement has been shown using using constrained Maximum Likelihood Linear Regression (cMLLR) model adaptation methods, while a fast adaptation is guaranteed by using linear Vocal Tract Length Normalization (lVTLN).We have achieved a relative gain of approximately 9.44% in the word error rate with unsupervised cMLLR adaptation. We also compare our ASR system with the Google ASR and show that, using adaptation methods, we exceed its performance.