Most automatic speech recognition systems are based on statistical models thatrequire training. While these types of systems have reached recognition ratesthat are sufficient for many purposes, they perform poorly for speaker typesthat are not present in the training material. Children are often absent fromtraining material for speech recognizers, and creating good training materialfor children can be difficult and expensive.To address this issue, this thesis focuses on using adult training material totrain a recognizer for children by adapting the training material duringtraining. Instead of performing speaker-dependent adaptation duringrecognition, where computational power may be scarce, and responsiveness may beessential, adaptation is performed during training towards a class of speakers.Using a combination of vocal tract length normalization (VTLN) and cepstralmean normalization during training, promising results have been obtained. In aconnected-digits task, a reduction in errors as high as 70% was shown, with areduction of almost 50% in a large vocabulary task. Using VTLN to warp thesame training material several times, combining these warped materials to trainone recognizer, a similar reduction in errors was shown, but with an increasedrobustness indicating a less speaker-dependent system. It is also shown that apiecewise linear warping method is better suited to warp adult speech to childspeech, than a bilinear warping method.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:ntnu-13646 |
| Date | January 2011 |
| Creators | Fjær, Bjørnar Grip |
| Publisher | Norges teknisk-naturvitenskapelige universitet, Institutt for elektronikk og telekommunikasjon, Institutt for elektronikk og telekommunikasjon |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0024 seconds