Data Selection Methods for Semi-Supervised Learning in Automatic Speech Recognition

For an automatic speech recognition (ASR) system to perform well on new target data, it must be trained on data from the same domain as the target. Semi-supervised learning, which integrates both manually transcribed and untranscribed data, can be beneficial in this context. However, manually transcribed data are often unavailable. This challenge can be addressed by generating a training dataset from a data pool whenever new target data are introduced.

By employing the methods proposed in this thesis, training data can be effectively selected from a multi-domain untranscribed data pool, leading to improved ASR performance. The findings of this research demonstrate the potential for leveraging large amounts of untranscribed data from multiple sources to enhance ASR systems.