Free Speech-to-Text Software for Serbian: A Step-by-Step Tutorial and Evaluation Using Noisy Interview Recordings

Read full text Summarise with VIRAT
Authors: Nevena Mijatović
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This technical note has two contributions. First, we introduce a free, cloud-based solution for automatic speech recognition (ASR) in Serbian, accompanied by a short step-by-step video tutorial (https://bit.ly/3KTHx8K). The solution uses OpenAI’s ASR called Whisper, the strongest model (called large-v3), which is otherwise typically accessible only to researchers with programming skills. Second, we evaluate this free Whisper-based solution in a common social science use case: transcription of interviews professionally recorded in noisy environments. Using an experimental design, we compared (1) the free Whisper solution, (2) a paid ASR service by ElevenLabs called Scribe (paid approx. $5 per 1-hour interview), and (3) manual transcription. Audio stimuli (N = 9) consisted of 1% randomly sampled segments from nine interviews. Dependent variables were: (a) transcription accuracy, quantified as word error rate (WER) by comparing ASR output to the same output corrected to full accuracy; (b) the proportion of meaning-altering errors (e.g., plan instead of plenum) among all errors; and (c) the time required to obtain a fully accurate transcript, operationalized as the ratio of transcription/correction duration to audio duration. One researcher manually corrected both ASR outputs and produced the manual transcript. Whisper produced a higher WER than ElevenLabs across the sample of interviews, t(8) = 5.47, p < .001 (Whisper: M = 0.204, SD = 0.082; ElevenLabs: M = 0.062, SD = 0.025). The proportion of meaning-altering errors did not differ reliably between systems, t(8) = 1.35, p = .107 (Whisper: M = 0.497, SD = 0.164; ElevenLabs: M = 0.384, SD = 0.328). Time requirements differed substantially across conditions, F(2,16) = 86.64, p < .001 , η² = .915, with all pairwise comparisons significant (p ≤ .01). Manual transcription required the most time (M = 6.11, SD = 1.80), followed by Whisper correction (M = 2.47, SD = 0.91), while ElevenLabs required the least time (M = 1.31, SD = 0.50). Overall, this evaluation suggests that the free Whisper-based solution can be a useful option for researchers who cannot afford paid ASR services. Although Whisper was less accurate than ElevenLabs in terms of WER, both systems produced meaning-altering errors, indicating that human inspection against audio remains necessary. Nevertheless, post-correction of machine output is likely to save substantial time compared to full manual transcription.

Fields of Study & Subjects

Speech Recognition and Synthesis Speech and Audio Processing Speech and dialogue systems

Publication details

Journal / source
Repository of the Institute for Philosophy and Social Theory (University of Belgrade, Institute for Philosophy and Social Theory)
Publication date
1 January 2026
Language
EN