A novel speaker adaptation approach for continuous densities HMM´s

An approach for speaker adaptation aiming to get high recognition performance from an HMM speech recognizer after a short training session with a new speaker is presented. The technique presented exploits the Gaussian multivariate nature of continuous density HMM distributions, to adapt the model parameters. This adaptation technique was applied to a 20-word vocabulary. It was tested on 70 new speakers after a training session of 2 to 5 repetitions of each vocabulary word. The experiments carried out have shown a significant improvement in the recognition performance, even when only two training tokens from the new speaker are used