Whisper-Based Lecture Transcription for Montenegrin Higher Education with a Practical Evaluation

October 2026
Conference: Smart Algorithms & Models: Your Reach to Advanced DialoguesAt: Seville, Spain
Stevan CakicTomo PopovicSrdjan KrcoIgor Culafic

Automatic speech recognition (ASR) is becoming part of the learning infrastructure in higher education, used to caption lectures and to turn spoken material into study script. It is reliable enough for routine classroom use in well-supported languages, but far less certain in smaller ones. Montenegrin is a representative case. It is not among the languages OpenAI Whisper supports, so an operator must fall back on a related South Slavic tag, and the consequences of that choice have not been measured systematically. This paper reports a quantitative study of which tag and which Whisper variant to use. The corpus is four university courses in computer science, 30 lectures and 37.9 hours of audio. Three Whisper variants (medium, turbo, large) are crossed with three tags (Croatian, Bosnian, Serbian), and every configuration is transcribed three times at each of six window lengths, giving 4,860 transcripts, of which 810 are at the four-minute window used for the main comparison.