DFCRN:煤矿场景下基于多尺度复数域的语音增强模型

DFCRN: multi-scale complex-domain speech enhancement model for coal mine environments

  • 摘要: 针对煤矿井下综采、综掘等强噪声工业场景中,强机械噪声、多源非平稳干扰与传输失真导致生产一线语音质量下降的问题,提出一种基于多尺度复数域的语音增强模型——扩张频率卷积循环网络(DFCRN)。该模型在传统卷积循环网络框架上进行2项结构改进:① 在卷积编码−解码结构中引入频率扩张复数卷积块(FDC Block),通过在频率维度采用分层递增的扩张率多尺度扩大频率感受野,以增强对不同频带特征及跨频带依赖关系的建模能力。② 构建堆叠的复数扩张前馈序列记忆网络(CD−FSMN),在复数域并行建模长程时序依赖与频谱记忆,保留实部与虚部的耦合关系以增强相位恢复能力,以改善复杂噪声条件下的语音特征恢复能力。在煤矿语音数据集Coal−single和公开数据集VoiceBank+DEMAND与DNS−2020上开展实验,结果表明DFCRN较频率递归卷积循环网络(FRCRN)、分组时序卷积循环网络(GTCRN)等对比模型具有更优的语音增强综合性能,在Coal−single数据集上DFCRN增强语音的语音混响调制能量比(SRMR)从原始含噪语音的7.695提升至10.723,深度噪声抑制平均意见分(DNSMOS)中OVRL由2.430提升至2.747,主观感知质量得分更高,验证了该模型能有效抑制真实煤矿复杂噪声,显著提升语音的清晰度与可懂度。

     

    Abstract: To address the degradation of frontline speech quality caused by intense mechanical noise, multisource non-stationary interference, and transmission distortion in noisy underground coal mine environments, such as fully mechanized mining and roadway development, a multi-scale complex-domain speech enhancement model—the Dilated Frequency Convolutional Recurrent Network (DFCRN) was proposed. Two structural improvements were made to the conventional convolutional recurrent network framework. ① A Frequency-Dilated Complex Convolutional Block (FDC Block) was introduced into the convolutional encoder-decoder structure. Progressively increasing dilation rates along the frequency dimension expanded the frequency receptive field at multiple scales, improving the modeling of features in different frequency bands and cross-band dependencies. ② A stacked Complex Dilated Feedforward Sequential Memory Network (CD-FSMN) was constructed to model long-range temporal dependencies and spectral memory in parallel in the complex domain. The coupling between real and imaginary components was preserved to improve phase recovery and thereby enhance speech feature recovery under complex noise conditions. Experiments were conducted on the Coal-single coal mine speech dataset and the public VoiceBank+DEMAND and DNS-2020 datasets. The results showed that DFCRN achieved better overall speech enhancement performance than comparative models, including the Frequency Recurrent Convolutional Recurrent Network (FRCRN) and the Grouped Temporal Convolutional Recurrent Network (GTCRN). On Coal-single, the Speech-to-Reverberation Modulation Energy Ratio (SRMR) increased from 7.695 for the original noisy speech to 10.723 for DFCRN-enhanced speech, and the Overall Quality (OVRL) score in the Deep Noise Suppression Mean Opinion Score (DNSMOS) increased from 2.430 to 2.747. Subjective perceptual quality scores were also higher. These findings confirm that the model effectively suppresses complex noise in real coal mines and significantly improves speech clarity and intelligibility.

     

/

返回文章
返回