Szczegóły publikacji
Opis bibliograficzny
Prediction and causality visualization in speech recognition / Adam Wróbel, Mikołaj Jarosławski, Paweł JEMIOŁO // W: Theory and engineering of dependable computer systems and networks : proceedings of the sixteenth international conference on Dependability of Computer Systems DepCoS-RELCOMEX : June 28 – July 2, 2021, Wrocław, Poland / eds. Wojciech Zamojski, [et al.]. — Cham : Springer Nature Switzerland AG, cop. 2021. — (Advances in Intelligent Systems and Computing ; ISSN 2194-5357 ; vol. 1389). — ISBN: 978-3-030-76772-3; e-ISBN: 978-3-030-76773-0. — S. 477–486. — Bibliogr., Abstr. — Publikacja dostępna online od: 2021-05-27
Autorzy (3)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 134385 |
|---|---|
| Data dodania do BaDAP | 2021-06-10 |
| DOI | 10.1007/978-3-030-76773-0_46 |
| Rok publikacji | 2021 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Springer |
| Konferencja | International Conference on Dependability of Computer Systems 2021 |
| Czasopismo/seria | Advances in Intelligent Systems and Computing |
Abstract
Speech Recognition is a rapidly developing aspect of modern Deep Learning research. Many up-to-date technologies, such as Google Assistant or Alexa, apply it. They change the way people interact with everyday devices, i.e., phones, cars. Despite being so influential, current AI models are complex black-boxes, and few people comprehend what lies beyond them. This paper aims to feature a Deep Learning architecture for Speech Recognition and visualizations of the model state when making the predictions. We incorporated class activation maps and the attention mechanism to increase the accuracy of the presented algorithm. The introduced model can analyze audio sequences of various lengths. Preliminary results suggest that the proposed method provides a satisfactory outcome. It can be successfully applied to get insight into the state of the model during predictions. This work might, therefore, help Data Scientists detect flaws, such as overfitting in their algorithms.
