基于高密度肌电的对称位置发音肌肉对语音识别贡献的研究

说话是人类正常生活中最重要的技能之一，是发音相关肌肉在神经中枢的控制下协调运动的结果。表面肌电图法(Surface Electromyography，sEMG)是目前采集肌肉电信号的常用方法，能检测到可靠的肌肉电生理信息。用肌电信号进行语音分类时，所选的电极位置对分类精度有重大作用。但目前基于 sEMG 的语音识别方法选取电极位置及数量时没有一个客观的指标，也不清楚发音相关的面、颈部左右两侧对称位置电极对肌电语音识别的贡献是否冗余。该文使用 120 通道电极(关于面中、颈中对称)采集了 8 名发音正常的受试者分别发 5 个中文单词和 5 个英文单词时的面、颈部 sEMG，考察了面、颈部左右两侧对称位置 sEMG 对语音识别的贡献。结果表明，发音过程中面、颈部左右两侧肌肉活动有相似的变化规律，但整个活动过程中面部对称位置的相关性比颈部低；使用颈部左侧、右侧的肌电信号进行语音分类的分类精度区别不大，而使用面部左、右两侧肌电信号的分类结果差异较明显。因此，颈部对称位置的 sEMG 信号对语音识别贡献程度具有一致性，而面部则不具有，这为后续研究减少电极数量和选择最佳通道提供了新思路。

The Study on the Left/Right Contributions of Articulatory Muscles inSpeech Recognition Using High-Density Surface Electromyography

Speech is one of the most important skills in human normal life. It is the result of the coordinatedmovement of the articulation-related muscles under the control of central cervous system. Surfaceelectromyography (sEMG) is a commonly used method for collecting electrical signals of muscles, whichcan detect reliable electrophysiological information. When using electromyographic signals on speechclassification, the selected electrode position plays an important role in classification accuracy. However, thecurrent sEMG-based speech recognition method does not have an objective index for selecting the positionand number of electrodes, and it is still unclear whether the contribution of the articulation related symmetricalposition electrodes on the left and right sides of the face and neck to speech recognition is redundant. Inthis study, the facial and neck sEMG of 8 subjects with normal pronunciation were collected by using a120-channel electrode (about facial and neck symmetry) when they pronounced 5 Chinese words and 5 Englishwords respectively. The contribution of sEMG in the symmetrical position of left and right sides of facial andneck to speech recognition was investigated. The results show that the muscles of the left and right sides of theface and neck had similar variation, but the correlation between the symmetrical positions of the face and neckwas lower than that of the neck. There was little difference in classification accuracy between the left and rightsEMG signals of the neck, but significant difference between the left and right SEMG signals of the face. Thus,sEMG signals from symmetrical positions in the neck are consistent in their contribution to speech recognition,whereas facial signals are not, which might provide useful clue to reduce the electrode number and select theoptimal location of channels for speech recognition.