Code-switching event detection based on delta-BIC using phonetic eigenvoice models

This paper presents a new paradigm for code-switching event detection based on delta Bayesian Information Criterion (ΔBIC). First, an automatic speech recognizer (ASR) and an articulatory feature (AF) detector are constructed. The intersyllable boundaries obtained from the ASR are regarded as the potential code-switching boundaries. To estimate the language likelihood, eigenvoice models (EVMs) are employed to model the relationship between the senones/articulatory attributes and their corresponding eigenvoices constructed from the training data for different languages. The Euclidean distance and the inner product-based direction between the eigenvoice vector of the input sentence and the eigenvoice vector of a senone or an articulatory attribute in the EVMs for different languages are calculated for ΔBIC-based language likelihood estimation. Then, an n syllable Bayesian mask centered at each potential boundary is then employed to output the likelihood of language change for the potential boundary. Finally, the dynamic programming algorithm is employed to search the best language sequence given the inter-syllable boundaries from the ASR. The proposed approach was evaluated on a Chinese-English codeswitching speech database and the results show that 71:93% accuracy for code-switching event detection can be obtained.