Rapid speech recognizer adaptation to new speakers

Discounted likelihood linear regression for rapid speaker adaptation

“…The DLLR adaptation schemes described above have been evaluated on a rapid adaptation task defined on the Switchboard corpus at the 1998 JHU LVCSR Workshop (Digalakis et al, 1999). Since this is an unsupervised adaptation task, the unadapted acoustic models are used to generate the transcriptions used for adaptation.…”

Maximum a posteriori adaptation of HMM parameters based on speaker space projection

“…Hence, in order to successfully develop ASR applications in real environments, it is crucial to take into account such mismatches. Since the information provided by a new speaker is very little in many applications, a variety of rapid adaptation techniques which use only a small amount of adaptation data are required (Digalakis et al, 1999;Kuhn et al, 2001b). Recently, many model based adaptation algorithms which modify hidden Markov model (HMM) parameters so as to better fit to a new speaker have been developed (Lee and Huo, 2000;Woodland, 2001;Sagayama et al, 2001).…”

Rapid speaker adaptation using probabilistic principal component analysis

“…Recently, there has been increasing interest in speaker adaptation techniques that require only a small amount of data from the target speaker. Such rapid adaptation schemes have been developed for modeling the dependencies between different speech units for effective use of a small adaptation data [2]. To estimate the dependencies between diverse units of speech, a large corpus of training speakers and a variety of correlation modeling approaches are used [2].…”