Skoči na glavni sadržaj

Izvorni znanstveni članak

https://doi.org/10.20532/cit.2026.1006584

Optimization and Development of Edge-Embedded Speech Recognition Module of Improved DFCNN-CTC

Fang Nie ; Chengdu Jincheng University, Chengdu, Sichuan Province, China
Lijuan Zheng ; School of Artificial Intelligence and Electronic Engineering, Sichuan Technology and Business University, Chengdu, Sichuan Province, China *

* Dopisni autor.


Puni tekst: engleski pdf 2.979 Kb

str. 203-218

preuzimanja: 0

citiraj


Sažetak

In view of the challenges faced by edge-device speech recognition of smart toys, such as limited computing resources, environmental noise interference, and specificity of children's pronunciation, this study proposes
a new lightweight speech recognition model. The study uses original spectrograms as input and introduces transformers to improve the deep full-sequence convolutional neural network-connected temporal classification model to achieve high-precision end-to-end recognition. This model uses dual-student difference structure mutual learning, sequence-level distillation, and KL imitation loss to narrow the performance gap between the student model and the teacher model. Experiments demonstrated that the improved model's word error rate on the children's speech data set reached 8.2%, which was 3.9% ~ 6.1% lower than Conformer, Branchformer, and other models. After the lightweight improvement, the model size was reduced from 125MB to 24MB, with an increase in WER of only 0.7%. After further compression, the model size was reduced to 8.1 MB, the inference delay was reduced to 155ms, the average power consumption was 45mW, and the energy efficiency ratio was improved by 4.4 times. In the home environment test, the wakeup success rate was 96.5%, the command recognition accuracy rate was 93.2%, and the false wake-up rate was 0.8 times/hour. The proposed method can improve the speech recognition efficiency of smart toys in edge environments and reduce their energy consumption.

Ključne riječi

smart toys; edge-embedding; speech recognition; DFCNN-CTC; mutual learning of knowledge distillation

Hrčak ID:

353530

URI

https://hrcak.srce.hr/353530

Datum izdavanja:

17.11.2026.

Posjeta: 0 *