Language Model Adaptation for Transcription of Banking Protocols

Authors: Ivan Kraljevski, Diane Hirschfeld

Abstract:

This paper presents an approach for adaptation of a LVCSR system on a specific domain - speech transcriptions for automated protocol generation during investment consultations. Because of the small amount of available domain-specific speech and textual data, it is not possible to create reliable statistical language model, therefore, word categories containing synonyms were used to train a word-class based model. To provide an appropriate domain-specific textual corpus for language model training, data augmentation was employed by creation of grammar rules and generation of large number of “artificial” sentences. Such language model could be used as standalone or could be merged with the general model. Recognition performance was compared across different language models: the domain-specific model, the general purpose model and as well as their weighted combinations. The results justified the proposed approach for domain-specific language modeling on banking protocols transcriptions.