Automatic Emergency Detection in Naval VHF Transmissions
Publicerad
Författare
Typ
Examensarbete för masterexamen
Program
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Abstract
As the proficiency of Speech-To-Text (STT) services increases, so does the possible
applications. This thesis explores the use of such services in a very special domain,
naval VHF transmissions. It explores STT service performance and details the development
of a domain-specific Speech-To-Text model based on the self-supervised
wav2vec 2.0 architecture. This enabled the recognition of emergency messages using
keyword detection and also created a foundation for more advanced intent analysis
in the future. The developed model outperforms Google on the naval domain and
achieves good classification results using keyword detection, managing to discern
most messages containing one or more keywords. This performance meant that
the model could be used as an aid for actual emergency message detection by Sjöfartsverket.
The research also shows that many of the pre-trained models do not
have adequate performance on the intended domain, but it was noted that using
semi-supervised methods such pre-trained models can be tuned to reach acceptable
performance levels. This can be done with smaller sets of domain-specific data to
achieve good results on the specific domains without the need for a completely new
model for each domain.
Beskrivning
Ämne/nyckelord
Automatic Speech Recognition, Speech-To-Text, Intent Analysis, Selfsupervised, wav2vec 2.0, Naval Environment, Emergency Messages