# Decentralizing Feature Extraction with Quantum Convolutional Neural Network for Automatic Speech Recognition

**Authors:** Chao-Han Huck Yang, Jun Qi, Samuel Yen-Chi Chen, Pin-Yu Chen, Sabato Marco Siniscalchi, Xiaoli Ma, Chin-Hui Lee  
**arXiv:** [2010.13309](https://arxiv.org/abs/2010.13309)  
**Date:** 2020-10-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.13309v2/

## Key points

- Decentralized approach for feature extraction in federated learning to address privacy-preservation concerns in speech recognition
- Utilizes a quantum convolutional neural network (QCNN) with a quantum circuit encoder and recurrent neural network (RNN) based end-to-end acoustic model (AM)
- Input speech is sent to a quantum computing server for Mel-spectrogram extraction
- Convolutional features are encoded using a quantum circuit algorithm with random parameters
- Encoded features are sent back to the local RNN model for final recognition
- Leverages advancements in quantum learning for secure models and prevention of privacy leakage attacks
- QCNN encoder achieves an accuracy of 95.12% in the decentralized model on the Google Speech Commands Dataset, outperforming previous architectures
- Analysis of different quantum circuit encoder architectures provides insights into designing QCNN-based feature extractors
- Correlation between proposed QCNN features, class activation maps, and input spectrograms revealed through neural saliency analyses
- Research contributes to addressing privacy concerns in speech recognition by combining quantum computing techniques with deep learning models

## Summary

This paper presents a novel decentralized approach for feature extraction in federated learning to address privacy-preservation concerns in speech recognition. The proposed approach utilizes a quantum convolutional neural network (QCNN) consisting of a quantum circuit encoder for feature extraction and a recurrent neural network (RNN) based end-to-end acoustic model (AM). To enhance model parameter protection in a decentralized architecture, the input speech is first sent to a quantum computing server where the Mel-spectrogram is extracted. The corresponding convolutional features are then encoded using a quantum circuit algorithm with random parameters. These encoded features are subsequently sent back to the local RNN model for final recognition. The decentralized framework proposed in this study leverages the advancements in quantum learning to ensure secure models and prevent privacy leakage attacks. Experimental results on the Google Speech Commands Dataset demonstrate that the QCNN encoder achieves an impressive accuracy of 95.12% in the decentralized model, outperforming previous architectures that employ centralized RNN models with convolutional features. Additionally, this paper conducts an extensive analysis of different quantum circuit encoder architectures to provide insights into designing QCNN-based feature extractors. Neural saliency analyses reveal a correlation between the proposed QCNN features, class activation maps, and input spectrograms. Overall, this research contributes to addressing privacy concerns in speech recognition by introducing a decentralized feature extraction approach that combines quantum computing techniques with deep learning models. The provided implementation serves as a valuable resource for future studies in this domain.