Phonotactic Information Based Language Identification

As an important research direction of Intelligent speech processing,automatic language identification can be defined as the process of determining which language a given utterance belongs to . With the acceleration of the globalization, people all over the world communicate with each other more and more frequently, and overcoming language barrier becomes very urgent. So, language identification is of great importance in many applications such as front-end of multilingual speech recognition, information retrieval, military intelligence gathering and public security, and has received widespread attention in the corresponding research fields.According to the different features used, the state-of-the-art language identification system falls into two main categories: acoustic information based language identification and phonotactic information based language identification. In the phonotactic information based language identification system, the given utterance is first transformed into a phone sequence with a phone recognizer, and then the language ID can be identified by the different phonotactic constraints between different languages. The phonotactic approach has attracted increasing attention for its top performance and good expansibility.This thesis concentrates on the study of phonotactic information based language recognition. After the construction of a complete language recognition system, which includes the phone recognizer and language models, we make progress on performance improvement and computational complexity reduction. The detailed works are as follows:Firstly, to select appropriate amount of data to achieve a phone recognizer training,we propose a method based on phone balance criterion, which picks phone balanced utterances from large amounts of speech data.Secondly, it is observed that for certain utterance, the noise in token sequence output from the phone recognizer is introduced due to the channel, speaker and background clutters. To address this problem, in this paper we propose to reduce the noise using the factor analysis: first, represent each utterance in N-Grams vector, then in this vector space, the factor analysis is applied to model the noise subspace, which will be reduced in final modeling process.Thirdly, to reduce the computational complexity and the redundancy of the feature vectors in the phone recognition followed by support vector machine strategy,, we carry out a novel approach for modeling higher order N-Grams using low order discriminative N-Grams, which get an improvement for language identification.