prikaz prve stranice dokumenta Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model
  Download
PDF 1.77 MB
master's thesis
Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model
Zagreb: University of Zagreb, Faculty of Humanities and Social Sciences, 2024. urn:nbn:hr:131:019095
Porupski, Ivan
University of Zagreb
Faculty of Humanities and Social Sciences
Department of Linguistics

Institutional repository: ODRAZ

Cite this document

Porupski, I. (2024). Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model (Master's thesis). Zagreb: University of Zagreb, Faculty of Humanities and Social Sciences. Retrieved from https://urn.nsk.hr/urn:nbn:hr:131:019095

Porupski, Ivan. "Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model." Master's thesis, University of Zagreb, Faculty of Humanities and Social Sciences, 2024. https://urn.nsk.hr/urn:nbn:hr:131:019095

Porupski, Ivan. "Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model." Master's thesis, University of Zagreb, Faculty of Humanities and Social Sciences, 2024. https://urn.nsk.hr/urn:nbn:hr:131:019095

Porupski, I. (2024). 'Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model', Master's thesis, University of Zagreb, Faculty of Humanities and Social Sciences, accessed 20 November 2024, https://urn.nsk.hr/urn:nbn:hr:131:019095

Porupski I. Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model [Master's thesis]. Zagreb: University of Zagreb, Faculty of Humanities and Social Sciences; 2024 [cited 2024 November 20] Available at: https://urn.nsk.hr/urn:nbn:hr:131:019095

I. Porupski, "Vrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model", Master's thesis, University of Zagreb, Faculty of Humanities and Social Sciences, Zagreb, 2024. Available at: https://urn.nsk.hr/urn:nbn:hr:131:019095

Metadata
TitleVrednovanje jezičnog alata za transkripciju govora: ParlaSpeech-HR ASR model
Title (english)Evaluation of a Language Tool for Speech Transcription: ParlaSpeech-HR ASR Model
AuthorIvan Porupski
MentorBožo Bekavac (mentor)
Committee memberVanja Štefanec (predsjednik povjerenstva)
Committee memberMarko Tadić (član povjerenstva)
Committee memberBožo Bekavac (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Humanities and Social Sciences
(Department of Linguistics)
Zagreb
Defense date and country2024-09-16, Croatia
Scientific / art field,
discipline and subdiscipline		HUMANISTIC SCIENCES
Philology
General Linguistics
Abstract
Ovaj rad fokusira se na vrednovanje jezičnog alata za transkripciju govora koristeći ParlaSpeech-HR ASR model i korpus. ParlaSpeech-HR korpus sadrži 404 tisuće snimki govora iz hrvatskog parlamenta, zajedno s njihovim transkriptima. U radu je korišten wav2vec2 akustični model, dok je kenLM (5-gramski) jezični model izrađen na temelju tekstualnog dijela korpusa. Spojeni su akustični i jezični model kako bi se stvorilo cjelovito ASR (Automatic Speech Recognition) rješenje. Vrednovanje je provedeno na uzorku od 1206 snimki iz govornog korpusa, koristeći metrike WER (Word Error Rate), CER (Character Error Rate) i F1 mjeru. Rezultati su pokazali da kombinirani akustični i jezični model postiže bolje performanse u usporedbi s modelom koji koristi samo akustičke podatke.
Abstract (english)
This paper focuses on the evaluation of a language tool for speech transcription using the ParlaSpeech-HR ASR model and corpus. The ParlaSpeech-HR corpus contains 404 thousand speech recordings from the Croatian parliament, along with their transcripts. The study utilizes the wav2vec2 acoustic model, while a kenLM (5-gram) language model was developed based on the textual part of the corpus. The acoustic and language models were combined to create a comprehensive ASR (Automatic Speech Recognition) solution. Evaluation was performed on a sample of 1206 recordings from the speech corpus, using WER (Word Error Rate), CER (Character Error Rate), and F1 metrics. The results indicate that the combined acoustic and language model achieves better performance compared to the model that uses only acoustic data.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:131:019095
Study programmeTitle: Linguistics (double major); specializations in: General Linguistics, Comparative Linguistics, Computational Linguistics, Aplied Linguistics and Cognitive Linguistics, Cognitive
Course: Computational Linguistics
Study programme type: university
Study level: graduate
Academic / professional title: magistar/magistra lingvistike (magistar/magistra lingvistike)
Type of resourceText
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2024-09-22 13:12:22