DOI: 10.11118/ecos2026_012

A TOOL FOR AUTOMATICALLY TRANSCRIBING IN-TERVIEWS WITH SPEAKER RECOGNITION

Lukáš Kleveta¹, Jaromír Landa¹

¹ Department of Informatics, Faculty of Business and Economics, Mendel University in Brno, Zemědělská 1, 613 00 Brno, Czech Republic

The present paper presents the design and implementation of a software tool for automatically transcribing audio recordings, supplemented by speaker recognition (diarization). The tool integrates two open-source components: The implementation of OpenAI’s Whisper AI model for speech-to-text conversion and the pyannote.audio library-or, alternatively, NVIDIA NeMo - for speaker recognition. The architecture of the system is implemented as a linear pipeline with modules for audio preprocessing, diarization, transcription, and result merging. It is possible to modify key processing parameters (model, language, thresholds, clustering method) via a configuration file without modifying the code. The functionality was verified on two test sets. The first was a short English recording with a single speaker. In this case, a quantitative comparison of Whisper AI models was made using the Levenshtein metric. The second was a one-minute recording with two speakers. In this case, a comparison was made of twenty configurations in terms of transcription, diarization, and computational complexity. The medium and large-v2 models demonstrated the highest degree of mutual agreement among transcriptions (97.79%), while the lowest was exhibited by the tiny and large-v3 models (82.88%). With regard to overall balance, the configuration employing the large-v2 model and the NeMo diarization module with spectral clustering was identified as the optimal solution. The findings substantiate the practical viability of open-source instruments for automated transcription. However, they concurrently underscore the constraints associated with multilingual recordings, the computational intricacy, and the susceptibility of diarization to variations in audio quality.

Klíčová slova: automatic speech transcription, speaker diarization, Whisper, pyannote.audio, NVIDIA NeMo, audio processing, open-source, Levenshtein distance, pipeline

stránky: 13-13, online: 2026