Deep Neural Network Acoustic Model Baseline for Character-Level Transcription of Naturally Spoken Czech Language

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F60461373%3A22340%2F20%3A43921328" target="_blank" >RIV/60461373:22340/20:43921328 - isvavai.cz</a>

Výsledek na webu

<a href="https://www-scopus-com.ezproxy.vscht.cz/record/display.uri?eid=2-s2.0-85098194179&origin=resultslist&sort=plf-f&src=s&st1=&st2=&sid=456c730de23256c12af0d4a383bd2de8&sot=b&sdt=b&sl=128&s=TITLE-ABS-KEY+%28Deep+Neural+Network+Acoustic+Model+Baseline+for+Character-Level+Transcription+of+Naturally+Spoken+Czech+Language%29&relpos=0&citeCnt=0&searchTerm=" target="_blank" >https://www-scopus-com.ezproxy.vscht.cz/record/display.uri?eid=2-s2.0-85098194179&origin=resultslist&sort=plf-f&src=s&st1=&st2=&sid=456c730de23256c12af0d4a383bd2de8&sot=b&sdt=b&sl=128&s=TITLE-ABS-KEY+%28Deep+Neural+Network+Acoustic+Model+Baseline+for+Character-Level+Transcription+of+Naturally+Spoken+Czech+Language%29&relpos=0&citeCnt=0&searchTerm=</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-030-63322-6_14" target="_blank" >10.1007/978-3-030-63322-6_14</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Deep Neural Network Acoustic Model Baseline for Character-Level Transcription of Naturally Spoken Czech Language

Popis výsledku v původním jazyce

Heavy portion of previous work in automated speech recognition (ASR) conducts their research and provides results exclusively for English spoken language, for which there are many ready-to-use datasets publicly available. ASR research for other less widely utilized languages suffers from the lack of high-quality speech datasets and insufficient experimental results to compare with new research. In this article, we aim to remedy this problem for the Czech language by proposing Deep Neural Network-based Acoustic Model (AM) architecture as well as providing experimental results for character-level transcription of naturally spoken utterances. The AM architecture was developed by utilizing working solutions from modern English language ASR research and further tailored for better performance on Czech language datasets by conducting a comparative experimental study for several versions of the AM. The models were trained on up to 331 h of naturally spoken Czech language utterances from the PDTSC 1.0 and ORAL2013 datasets and validated on a 5.5-hour excerpt from the PDTSC 1.0 dataset. The results show that our final AM architecture can reach an average of 26.6 % transcript character error rate (CER) on the validation set when trained with all of the available training data. We believe that the final AM architecture presented in this paper and the experimental results can serve as a baseline for further Czech language ASR research.

Název v anglickém jazyce

Deep Neural Network Acoustic Model Baseline for Character-Level Transcription of Naturally Spoken Czech Language

Popis výsledku anglicky

Heavy portion of previous work in automated speech recognition (ASR) conducts their research and provides results exclusively for English spoken language, for which there are many ready-to-use datasets publicly available. ASR research for other less widely utilized languages suffers from the lack of high-quality speech datasets and insufficient experimental results to compare with new research. In this article, we aim to remedy this problem for the Czech language by proposing Deep Neural Network-based Acoustic Model (AM) architecture as well as providing experimental results for character-level transcription of naturally spoken utterances. The AM architecture was developed by utilizing working solutions from modern English language ASR research and further tailored for better performance on Czech language datasets by conducting a comparative experimental study for several versions of the AM. The models were trained on up to 331 h of naturally spoken Czech language utterances from the PDTSC 1.0 and ORAL2013 datasets and validated on a 5.5-hour excerpt from the PDTSC 1.0 dataset. The results show that our final AM architecture can reach an average of 26.6 % transcript character error rate (CER) on the validation set when trained with all of the available training data. We believe that the final AM architecture presented in this paper and the experimental results can serve as a baseline for further Czech language ASR research.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2020

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

4th Computational Methods in Systems and Software, CoMeSySo 2020

ISBN

978-3-030-63321-9

ISSN

2194-5357

e-ISSN

—

Počet stran výsledku

16

Strana od-do

170-185

Název nakladatele

Springer

Místo vydání

Heidelberg

Místo konání akce

Vsetín

Datum konání akce

14. 10. 2020

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—