Multilingual query-by-example spoken term detection in Indian languages

Spoken language processing poses to be a challenging task in multilingual and mixlingual scenario in linguistically diverse regions like Indian subcontinent. Common articulatory based framework is explored for the representation of phonemes of different languages. This framework is designed to handle typical features like aspirated plosives, nasalized vowels, combined letters, unvoiced retroflex plosive which are found in majority of Indian languages. It is trained with two languages. Different strategies like transfer training and joint training are studied to adapt English trained neural networks with smaller amount of Bangla data. It is observed that such training not only improves Query-by-Example Spoken Term Detection (QbE-STD) in the language of same language family like Hindi but also other Indian languages like Tamil and Telugu. While cross lingual adaptation of neural networks with a language specific softmax layer has been studied earlier in context of speech recognition, this work presents an architecture which is language independent uptil softmax layer. It is observed that this architecture has higher accuracy for unseen languages, is more compact and can be adapted more easily for new languages in comparison to the classic phoneme posteriorgrams based architecture.