基于RNN滤波的厨房环境语音降噪技术研究

Research on Speech Noise Reduction Technology for Kitchen Environment Based on RNN Filtering

本文提出了一种基于门控循环单元(GRU)的多层RNN滤波网络模型。通过训练RNN网络来学习含噪音频信号的时间相关性和局部模式,从而实现对实时语音的前端降噪处理。使用PESQ和STOI指标对滤波结果进行评价,结果显示相比传统的谱减法和Wiener法等滤波方法,基于RNN滤波模型的降噪技术在人耳感知效果方面更优。实验结果表明,该技术能够在降噪的同时保持语音的可识别度,减少厨房环境噪音对后续语音识别的影响。

This paper proposes a multi-layer RNN filtering network model based on Gated Recurrent Units(GRU).By training the RNN network to learn the temporal correlations and local patterns of noisy audio signals,the model achieves front-end noise reduction processing for real-time speech.The filtering results are evaluated using PESQ and STOI metrics,demonstrating that the RNN-based filtering model's noise reduction technique is superior in terms of auditory perception compared to traditional methods such as spectral subtraction and Wiener filtering.Experimental results show that this technology can maintain speech intelligibility while reducing noise,minimizing the impact of kitchen environment noise on subsequent speech recognition tasks.