Acoustic modeling for hindi speech recognition in low-resource settings

摘要

We propose an approach for acoustic modeling of Hindi speech by borrowing from English data, for the purpose of Hindi LVCSR. Hindi, like many Indian languages, has a significant speaker base but there have not been a lot of resources to obtain large amounts of transcribed Hindi data for LVCSR. We compare a baseline Gaussian model-sharing approach with DNN training. A widely used data-borrowing method with DNN is to firstly train a DNN with English, for which a large amount of training data is available; then the whole DNN, except the last layer, is fine-tuned by using the target Hindi data. We propose to do phonetic mapping between Hindi and English in the first stage, training Hindi acoustic models by sharing data between Hindi-English phone pairs in the second stage, and finally fine-tuning the acoustic model by using the Hindi data. We evaluate and compare these approaches with experiments using 1 hour of transcribed Hindi data and 15 hours of Wall Street Journal English data. Experiments show that the proposed method significantly outperforms conventional baseline models in a low-resource setting for phone recognition tasks.