【摘要】 在煤矿井下复杂噪声环境中，语音信号与噪声的频谱能量呈现非均匀分布特性，导致承载关键语义信息的低能量频段易被噪声掩盖，产生显著的频谱掩蔽效应。上述现象严重制约了语音识别的精度和鲁棒性。为此，提出一种敏感频段增强下的煤矿井下调度语音识别方法。通过等效矩形带宽滤波器组将原始语音分解为多个子带，对各个子带展开分帧，通过离散余弦变换转换到频域，整合全部增强子带，完成煤矿井下调度语音敏感频段增强。根据增强后调度语音特点，采用一维可分离卷积构造局部注意力模块，在频谱掩蔽效应下依然可对低能量语音信号进行局部特征提取，捕捉语音中关键语义信息，实现调度语音识别。实验研究表明，该方法能够显著降低频带特定失真指数(Band-Specific Distortion Index,BSDI)，在高噪声环境下保持较低的字错误率，任务完成率接近100%，验证了其在煤矿井下复杂环境中的有效性和鲁棒性。更多还原

【Abstract】 In the complex noisy environment underground in coal mines,the spectral energy of speech signals and noise exhibits a non-uniform distribution. This causes low-energy frequency bands carrying key semantic information to be easily masked by noise,resulting in a significant spectral masking effect. This phenomenon severely restricts the accuracy and robustness of speech recognition. To this end,a method for underground coal mine dispatch speech recognition based on sensitive frequency band enhancement is proposed. The original speech is decomposed into multiple subbands using an equivalent rectangular bandwidth filter bank. Each subband is then framed and converted to the frequency domain using a discrete cosine transform. All enhanced subbands are then integrated to enhance the sensitive frequency bands of underground coal mine dispatch speech. Based on the characteristics of the enhanced dispatch speech,a local attention module is constructed using a one-dimensional separable convolution. This module can extract local features from the low-energy speech signal even under the spectral masking effect,capture key semantic information in the speech,and achieve dispatch speech recognition. The experimental studies have shown that this method can significantly reduce the band-specific distortion index( BSDI),maintain a low word error rate in high-noise environments,and achieve a task completion rate close to 100%,verifying its effectiveness and robustness in the complex environment of underground coal mines.更多还原