Performance Analysis of Phoneme Based Voice Recognition System on Different Accents

This paper presents a performance analysis of a phoneme-based voice recognition system, focusing on the impact of different English accents and its implementation on resource-constrained Edge AI platforms. The study evaluates how phoneme similarities affect recognition accuracy, especially words with shared vowel or consonant structures across accents. Experiments were conducted on a microcontroller-based platform, optimized for low-power, low-memory constraint, using accent-specific voice data generated by text-to-speech engines. The platform utilizes Keyword Spotting (KWS) on-device, employing phenome based deep neural networks (DNNs) for real-time processing. Results show variance in recognition accuracy based on accent, with vowel sounds playing a critical role. Additionally, memory trade-offs were mentioned when introducing multiple phoneme sequences to improve performance in accent adaptation. Recommendations are made for optimizing voice recognition on Edge AI platforms, ensuring both accuracy and efficiency across diverse accents.