Voice Intelligence at the Brilliance Edge
Chu, Anh Jr (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
Hyväksymispäivämäärä
2025-12-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025122112038
https://urn.fi/URN:NBN:fi:tuni-2025122112038
Tiivistelmä
Voice assistants in cloud services such as Apple Siri, Amazon Alexa, Microsoft Cortana, and Google Assistant have been the main methodology for creating the voice assistant. With the basic components of Automatic Speech Recognition (ASR), Natural Language Processing (NLP), and Text to Speech (TTS), a cloud provider can produce a general voice assistant that serves most users. However, voice assistants in the cloud consume lots of energy and are highly dependent on internet connectivity. Therefore, this thesis proposes two solutions that involve deploying a voice assistant on an embedded device like Raspberry PI 5. Voice assistant in a resource-constrained device will consume less energy, incur less cost to the user, and can be used in rural areas where there is no internet connectivity. A specific market in a rural area can fully utilize the embedded voice assistant without losing its accuracy and latency.
The embedded solution can utilize LLM when there is a significant development in generative AI. But instead of using a large model, a memory-constrained device can only use a small model, which significantly reduces the size from the normal LLM. A small language model can easily target specific topics and fulfill users’ requirements without training on a large amount of dataset. The small model also vastly increases the generation speed, which is beneficial for resourceconstrained hardware.
In this thesis, a literature review is conducted to see what methodologies are used to create a small voice assistant in an embedded device. The thesis proposed 2 solutions: one utilizes a small language model and one utilizes a Question Answer Pair model. Both provided assistants also utilize RAG, which can retrieve information stored in its database before responding to the user. The thesis will conduct constructive research to design two prototypes and demonstrate how applicable the prototypes are in real-life situations. Two voice assistants were created during the research’s implementation phases, and a full comparison of both solutions is conducted. The result showed that a small language model solution always responds to user questions adaptively, while a Question Answering Pair provides deterministic answers. The thesis will later reflect the original theories of how the provided solutions can have more advantages compared to the cloud voice assistant in certain circumstances.
The embedded solution can utilize LLM when there is a significant development in generative AI. But instead of using a large model, a memory-constrained device can only use a small model, which significantly reduces the size from the normal LLM. A small language model can easily target specific topics and fulfill users’ requirements without training on a large amount of dataset. The small model also vastly increases the generation speed, which is beneficial for resourceconstrained hardware.
In this thesis, a literature review is conducted to see what methodologies are used to create a small voice assistant in an embedded device. The thesis proposed 2 solutions: one utilizes a small language model and one utilizes a Question Answer Pair model. Both provided assistants also utilize RAG, which can retrieve information stored in its database before responding to the user. The thesis will conduct constructive research to design two prototypes and demonstrate how applicable the prototypes are in real-life situations. Two voice assistants were created during the research’s implementation phases, and a full comparison of both solutions is conducted. The result showed that a small language model solution always responds to user questions adaptively, while a Question Answering Pair provides deterministic answers. The thesis will later reflect the original theories of how the provided solutions can have more advantages compared to the cloud voice assistant in certain circumstances.