Work

Optimering av Talarverifieringsmodeller för Resursbegränsade Enheter

Year: 2025
Type: dissertation
Abstract: This thesis explores the development and optimization of machine learning models for speaker verification on resource-constrained devices. The work addresses the challenge of balancing classification accuracy, inference speed, and memory usage. Speaker recognition is commonly implemented using cloud-based models, but these require a stable internet connection and introduce latency and privacy concerns. By moving inference to the edge, such limitations can be mitigated, though this presents new constraints in computational power and memory capacity. The project initially aimed to compare Convolutional Neural Network (CNN) and Random Forest (RF) classifiers using multiple preprocessing approaches, including Mel-Frequency Cepstral Coefficient (MFCC), spectrogram-based methods, and contrastive learning. Early experiments showed that RF models consistently underperformed in accuracy, inference speed, and memory efficiency compared to compact CNN architectures. As a result, the focus shifted toward scaled-down CNNs, in particular a Res2Netbased architecture. MFCC preprocessing was selected as the most effective feature extraction method for this use case. Models were trained and evaluated using both publicly available datasets and a custom dataset recorded from eight participants. The study also compared digital evaluation results with physical, on-device testing. Optimization techniques such as quantization and knowledge distillation were applied in attempt to further reduce model size and improve inference time, with trade-offs in accuracy. The compression methods used did not outperform the scaled-down version of the model. Notably, scaling down the Res2Net module reduced a model from 1700 KB to just 140 KB while increasing the Equal Error Rate (EER) by only around 1%. Results demonstrate that small CNN models can achieve high speaker recognition accuracy. Performance was found to be strongly influenced by the size and diversity of the training data, with a greater variety of impostor samples improving rejection performance. The findings contribute to the broader field of small-scale machine learning by showing that techniques developed for speech classification that could possibly be adapted to other forms of time-continuous data, offering opportunities for future applications in areas such as environmental monitoring and biomedical signal analysis.
Author: Lindberg, Axel
Language: English
Open Access status: green
OpenAlex ID: W7112078567
Cites: 41
Cited by:
Related to: 10