Multilingual acoustic modeling: a unified approach

Abstract

Spoken language processing technologies have matured enough to be used in commercial applications. Yet, services based on these technologies are not very popular in multilingual societies such as India, in spite of their wide applications. India is a multilingual society with more than 30 languages spoken across the country, and at least three languages spoken in most of the major cities. Further, people tend to mix words across languages, necessitating the use of multilingual solutions. Then, data collection for some of the languages like Tulu, a South Indian language, is extremely difficult if not impossible. People speaking the language are spread across the country, mixing with people speaking other major languages. This makes the language Tulu acoustically very diverse. Further, the total number of people speaking the language is less than a few million. The importance of a language cannot be related to the number of people speaking the language and could be even political at times. A multilingual solution would therefore be particularly attractive. When building multilingual acoustic models, language specific variations in the features increase and acoustic models becomes weak as a result. A possible solution is to capture these variations efficiently, in a way that it does not affect the performance of the system. Another alternative is to use robust features that are less sensitive to languages. Language identification is an integral part of any multilingual spoken language processing system. Language identification can help move the acoustic space of the language independent system to the subspace of the language. Or, it could be used to select the language specific acoustic model from many monolingual systems available. Phone recognition followed by language modeling, phonotactic approach, is one of the popular approaches to language identification for its simplicity and state-of-the-art performance.