Automatic speech-to-text transcription in an ecuadorian radio broadcast context

Abstract

A key element to enable the analysis and accessing to radio broadcast content is the development of automatic speech-to-text systems. The building of these systems has been possible given the current available of different speech resources, models, and open source services designed mainly for English language. However, the most of these tools have been migrated to other languages like Spanish for avoiding the creation of these systems from scratch. Despite existing efforts there is no clear evidence of the tools that can be used to convert audio to text in other dialects of Spanish. Also, the most of these systems are trained to consider a specific context, therefore, audio transcription systems personalized for a language and a specific context are needed. This article describes the implementation of an architecture oriented to automatic speech-to-text transcription applied on Ecuadorian radio broadcasters, using available free tools for performing audio segmentation and transcription. The selected tools were evaluated measuring their performance and facilities for adjusting to the defined architecture. At the end, a Web application was developed and its final performance was compared with IBM Watson speech to text service; the results show that the proposed system improves the accuracy and achieves a Word Error Rate around 10%. The obtained results allow to suggest the use of a free tools set in order to train models oriented to specific speech-to-text transcription scenarios.