An Edge–Cloud Collaborative Architecture for Real-Time Speech Recognition Using MobileNet and Conformer

In the modern era, voice recognition has gained significant attention and has become an essential component of smart devices and embedded systems, enabling human-machine interaction. However, integrating accurate voice recognition models on low-resource edge devices is challenging because of memory constraints and computational limitations. The existing end-to-end on the device speech recognition system based on MobileNetv3 and Connectionist Temporal Classification (CTC) decoder struggles in longer or unclear speech and does not involve any cloud architecture, which limits its capability to support more complex situations. To overcome these limitations, this study introduces an edge–cloud collaborative architecture for voice recognition. First, data from LibriSpeech were collected and pre-processed using log-Mel spectrogram extraction, normalization, and transcript alignment. This architecture comprises both edge and cloud devices. The edge server integrates the MobileNetv3-CTC model to handle short-or high-confidence speech, whereas the cloud server consists of a conformer model that handles uncertain or longer utterances. The proposed Edge-Cloud collaborative architecture achieved better results with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{4. 9 2 \%}$</tex> a Word Error Rate (WER) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{2. 8 7 \%}$</tex> and Character Error Rate (CER) while maintaining real-time performance.