Using language models to improve a speech recognition based maritime emergency call detection system

Abstract

Novel applications of the transformers architechture as well as the availability of pre-trained models have drastically reduced the amount of data required to train successful speech-to-text (STT) models. By using the Connectionist Temporal Classification (CTC) algorithm, the process is further simplified as the training data does not have to be pre-segmented. This work aims to improve the performance of such a model developed to detect maritime VHF radio emergency calls by adding a language model to the CTC-decoding. We experiment with language models trained on several different text corpora and apply language models both in the decoding and on the resulting transcripts. The results indicate the importance of large amounts of domain-specific text. The results also show that a reduced Word Error Rate (WER) does not necessarily lead to an improvement in contextual comprehension. Finally, it is shown that relatively large improvements are given by fine-tuning various pre-trained STT-models on a curated dataset.