Transfer learning for cross-lingual automatic speech recognition

In this study, an instance based transfer learning phoneme modeling approach is presented to mitigate the effects of limited data in a target language using data from richly resourced source languages. A maximum likelihood (ML) learning criterion is introduced to learn the model parameters of a given phoneme class using data from both the target and source languages. Each phoneme was modeled using a 3 state, 1 Gaussian mixture HMM. Turkish and English were chosen to be the target and source languages respectively. It was found that using only 20 utterances from Turkish, the monophone recognition accuracy in Turkish using transfer learned HMMs is close to the levels of accuracy achieved using standard HMMs when 100 or more utterances from the Turkish training corpus were used.