Exploring energy-efficient key word spotting on Android using network compression techniques and a model selection algorithm

ŠTIMEC, GAŠPER (author), Machidon, Octavian Mihai (mentor)

This thesis explores the development of an Android application for real-time keyword spotting, utilizing neural network based models deployed locally on the device. A central aspect of this work is the implementation of an algorithm that selects the most appropriate level of model complexity based on internal factors such as the device’s battery level and previous inference outcomes. Three models used were created using varying architecture designs and post-training quantization. The application captures audio through the device’s microphone, processes it to extract MFCC features, selects the model and performs classification, ensuring that the processing pipeline remains consistent with the model’s training conditions. Beyond the application itself, this thesis aims to provide insights into the performance of the adaptive complexity algorithm. It evaluates the trade-offs between energy consumption and classification accuracy across different app configurations: using a single model versus employing the adaptive algorithm.

• Language: English
• Keywords: keyword spotting, neural network, MFCC
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-164830
• COBISS.SI-ID: 214137859
• Publication date in RUL: 13.11.2024

Secondary language

• Language: Slovenian
• Title: Raziskovanje energetsko učinkovitega zaznavanja ključnih besed na Androidu z uporabo kompresijskih tehnik in algoritma za izbiro modela

Abstract

Ta diplomska naloga predstavlja izvedbo Android aplikacije, namenjene prepoznavanju ključnih besed v realnem času z uporabo TensorFlow Lite modelov, nameščenih lokalno na napravi. Poudarek je na implementaciji algoritma, ki izbere najprimernejši nivo kompleksnosti klasifijkacijskega modela na podlagi notranjih parametrov, kot so raven baterije naprave in prejšnji rezultati sklepanja. Omenjeni trije modeli so bili ustvarjeni z uporabo različnih arhitektur nevronskih mrež in s kvantizacijo. Aplikacija zajema zvok preko mikrofona naprave, ga obdela za pridobitev MFCC značilk, izbere model in nato izvaja klasifikacijo za prepoznavanje ključnih besed. Poleg same aplikacije, se teza posveča tudi evalvaciji delovanja algoritma za prilagodljivo kompleksnos in oceni kompromis med porabo energije in natančnostjo klasifikacije pri različnih konfiguracijah aplikacije: uporaba enega samega modela v primerjavi z uporabo algoritma.

• Keywords: zaznavanje ključnih besed, nevronska mreža, MFCC