A Case Study in Comparative Speech-to-Text Libraries for Use in Transcript Generation for Online Education Recordings

Abstract

With a proliferation of Cloud based Speech-to-Text services it can be difficult to decide where to start and how to make use of these technologies. These include the major Cloud providers as well as several Open Source Speech-to-Text projects available. We desired to investigate a sample of the available libraries and their attributes relating to the recording artifacts that are the by-product of Online Education. The fact that so many resources are available means that the computing and technical barriers for applying speech recognition algorithms have decreased to the point of being a non-factor in the decision to use Speech-to-Text services. New barriers such as price, compute time, and access to the services? source code (software freedom) can be factored into the decision of which platform to use. This case study provides a beginning to developing a test-suite and guide to compare Speech-to-Text libraries and their out-of-the-box accuracy. Our initial test suite employed two models: 1) a Cloud model employing AWS S3 using AWS Transcribe, 2) an on-premises Open Source model that relies on Mozilla's DeepSpeech[1]. We present our findings and recommendations based on the criteria discovered. In order to deliver this test-suite, we also conducted research into the latest web development technologies with emphasis on security. This was done to produce a reliable and secure development process and to provide open access to this proof of concept for further testing and development.