Improved methods for vocal tract normalization

Acoustic variability and automatic recognition of children’s speech

“…During both training and testing a grid search over 21 warping factors evenly distributed, with step 0.02, in the range 0.80-1.20, was performed. The training and recognition procedures adopted for implementing VTLN follow closely those proposed in (Welling et al, 1999) and are described in detail in (Giuliani et al, 2006).…”

Fast Search for Large Vocabulary Speech Recognition

“…However, the computation time is more than doubled. Using Ï instead of the actually spoken, but unknown transcription Ï does not degrade the recognition performance even if the preliminary transcription has a large word error rate in the order of 20 to 30% (Welling et al, 1999). …”

“…To estimate the warping factor of the test speaker without a preliminary recognition pass, Lee et al (1996) and Welling et al (1999) suggested a text-independent method using Gaussian mixture models. The approach used here relies on a separate emission distribution ÈÖ´ « µ for each warping factor «, where denotes the sequence of acoustic vectors and « denotes the «-dependent distribution parameters.…”

“…the vectors are normalized using the optimal speaker dependent warping factors (Welling et al, 1999).…”

The RWTH English lecture recognition system

“…Furthermore, vocal tract length normalization (VTLN) is applied to the MFCC features. The VTLN warping factors are obtained from a Gaussian classifier (fast-VTLN) [18]. In addition, we perform speaker adaptation with constrained maximum likelihood linear regression (CMLLR) [19].…”