Ääreanalüütika ressursikasutuse uuring kõne tekstiks muundamise näitel Android seadmes
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Igal aastal võetakse kasutusee miljardeid uusi nutikaid seadmeid. Osad neist on klassikalised eraldiseisvad seadmed, kuid enamik on ühendatud internetti või integreeritud nutiseadmete, autode ja muude tarkade masinate sisse peitu. Enamik seadmeid on ühendatud lokaalsete juhtseadmetega ja läbi juhtseadmete internetiga. Andmete hulk mida andurid igal aastal toodavad kasvab pidevalt ja võib lähiajal tekitada võrguühendustes ummikuid. Klassikaline isoleeritud tark-andur ja kesksesse pilve andmete saatmine ei ole enam alati parim lahendus seega viimastel aastatel kogub populaarsust lokaalne anduri andmete töötlus enne kesksesse pilve andmete saatmist. Samas enamik seadmete riistvara jõudlus on piiratud ja sega tuleb hoolsalt jälgida töötlemiseks vajalike ressursside olemasolu või kasutust. Töös uuritakse kolme kõne tekstiks muundamise lahendust ja iga lahenduse mälu, protsessori, võrgu ja energia kasutust. Testimiseks kasutatakse Sphinx ja Google kõnetuvastuse teeke. Magistritöö tulemusena näidatakse, et Android seadmes on võimalik edukalt taustal teha aktiivset kõnetuvastust ja suurim ressursikasutus on umbes 5\\% akut tunnis.
Billions of sensors are currently deployed around the world. Some are standalone sensors while others can be found in smart-phones, wearables, cars, machinery, buildings, street lights, wind turbines and other places too numerous to mention. These sensors are connected to intermediate edge devices which provides connectivity to the core network. The amount of data generated by sensors is staggering and with the rapid growth of sensors deployed, generated data will only continue to increase. The traditional way of handling data where generated data is sent to the network for analysis and decision made is already inefficient and completely impractical in most applications. A better approach would be to perform these analytics and decision making tasks on the edge devices. But due to the very limited available resources on edge devices, it is important to first analyze the computing resource utilization of sample applications running on edge device in order to understand what computational tasks are possible on these edge devices. This thesis aims to take Speech Recognition as a case study and analyze its resource consumption on an edge device. The thesis further aims to explore the possibility of implementing long running tasks on the edge without significantly impacting the limited edge resources. Finally, we investigate the possible impact of performing additional speech analytics tasks on the edge.
Billions of sensors are currently deployed around the world. Some are standalone sensors while others can be found in smart-phones, wearables, cars, machinery, buildings, street lights, wind turbines and other places too numerous to mention. These sensors are connected to intermediate edge devices which provides connectivity to the core network. The amount of data generated by sensors is staggering and with the rapid growth of sensors deployed, generated data will only continue to increase. The traditional way of handling data where generated data is sent to the network for analysis and decision made is already inefficient and completely impractical in most applications. A better approach would be to perform these analytics and decision making tasks on the edge devices. But due to the very limited available resources on edge devices, it is important to first analyze the computing resource utilization of sample applications running on edge device in order to understand what computational tasks are possible on these edge devices. This thesis aims to take Speech Recognition as a case study and analyze its resource consumption on an edge device. The thesis further aims to explore the possibility of implementing long running tasks on the edge without significantly impacting the limited edge resources. Finally, we investigate the possible impact of performing additional speech analytics tasks on the edge.