Institutional Repository of Vilnius University
Lietuvių kalbos atpažintuvų vertinimas /
Abstract
dc:descriptionThis work explores the development of ASR technologies and the main challenges of developing them and adapting them to the Lithuanian language. Additionally, it evaluates the accuracy and performance speed of different speech-to-text transcribers capable of recognizing Lithuanian speech and offering API integration capabilities. In this work the following transcribers are tested and evaluated: Google Cloud Speech-to-Text, Whisper (and it's Tiny, Base, Small, Medium, Large and Turbo models), Microsoft Azure Cognitive Services, Amazon Transcribe, AssemblyAI and Deepgram (Nova-2 native model and Whisper Large model integration). The used datasets consist of high-quality audio recordings with a length of up to 1 min., as well as longer recordings with a length of up to 10 min., which are characterized by lower quality, various recording settings and diversity of speaker genders. The transcribers were evaluated using WER, CER, MER, WIL and latency. The results showed that the most accurate transcribers were Microsoft Azure Cognitive Services, accordingly having 6.05% and 13.29% average WER, and Amazon Transcribe, accordingly having 6.54% and 11.87% average WER, depending on the used dataset. These transcribers also demonstrated fast performance speeds, making it difficult to claim that higher accuracy results in a slower performance. Furthermore, no correlation was noticed between recognition accuracy and the gender of the speaker in the audio recordings. The results of the most accurate transcribers were also compared to the results seen in a different work, which used the same dataset, showing some of the best WER and CER values seen in the comparison.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Strimaitis, Gustas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD229555837&prefLang=en_US