SCS Undergraduate Thesis Topics
|Hao Yee Chan||Roni Rosenfeld||Discriminative Pronunciation Learning for Speech Recognition for Resource Scarce Languages|
We develop pragmatic solutions that create small vocabulary speech recognizers at a fraction of the cost and time that current packages require. Little to no expertise in speech recognition is needed in training a recognizer for any language, and only a few audio samples per word are required. Hence, these solutions are ideal for targeting languages that have a small or economically disadvantaged user base which are typically ignored by the commercial world. In particular, building on previous work, we design algorithms that utilize current speech recognition engines to generate pronunciations which map phonemes across languages. Using discriminative training techniques, we further improve the pronunciation generated, resulting in higher recognition accuracy.