To address the issue of insufficient pathological speech feature extraction in existing research, which leads to low recognition rates for pathological speech, a pathological speech recognition algorithm based on a dual-branch residual network was proposed. Considering the complex and diverse speech symptoms of dysarthria patients, wideband and narrowband spectrograms were used as inputs to the network. An adaptive feature extraction residual block was introduced, which comprehensively captures pathological features from multiple dimensions, including position, channel, filter, and pixel, through full-dimensional dynamic pixel attention convolution. A dual-stream complementary fusion module was proposed, where the weighted fusion of features not only preserves critical information from each branch but also achieves optimal feature representation through cross-dimensional interactions, thereby improving the accuracy of pathological speech recognition. Experiments conducted on the Chinese pathological speech dataset THE-POSSD and the publicly available Western pathological speech dataset UA-Speech demonstrate the effectiveness and generalization capability of the proposed algorithm.
在早期的病理语音识别研究中,传统的机器学习方法占据主导地位。研究人员往往通过对输入的语音信号进行特征提取,计算声学特征,再利用分类器进行分类。Saska等[3]提取梅尔频率倒谱系数(Mel frequency cepstral coefficients,MFCC),深入研究不同帧长对病理语音检测的影响,采用支持向量机进行分类。梁正友等[4]提出一种基于多特征组合的病理语音识别方法,通过从韵律特征、频谱特征、人耳听觉特征、嗓音质量特征和声道模型特征的多特征组合中选择出分类准确率最高的特征子集,利用支持向量机进行分类。
FWA模块主要由空间通道重建卷积(spatial and channel reconstruction convolution, ScConv)[13]和高效多尺度注意力机制(efficient multi-scale attention,EMA)[14]组成,尽管宽带频谱图和窄带频谱图分别捕捉了语音信号的不同方面,这些特征终归来自同一条语音信号,因此在合并两个分支特征中不可避免地存在冗余信息。为了更好地融合这些特征并减少冗余,利用ScConv进行空间和通道重构,减少冗余信息,增强特征表达。ScConv模块由空间重构单元(spatial reconstruction unit, SRU)和通道重构单元(channel reconstruction unit, CRU)组成。SRU如图5所示,利用组归一化(group normalization)中的缩放因子来量化不同频带在空间维度上的特征显著性,接着通过sigmoid将这些缩放因子映射到[0,1]区间,结合阈值门控生成信息权重W1和非信息权重W2,这一过程有效识别出了包含更多有效信息的区域和相对冗余区域。随后,将两部分特征根据信息权重和非信息权重拆分为不同子块并进行交叉相加,通过元素级加法运算,将原本分散在不同频率和时域的特征片段进行整合,充分重构并融合局部特征信息,这样不仅消除冗余信息的干扰,还促进不同频带信息的相互补充,最后拼接重构出空间维度精炼后的特征。
所采用的病理语音数据集为由太原理工大学联合北京大学第一医院太原市中心医院共同录制的包含声学、声门和视频数据的山西徐沟地区卒中后言语功能障碍患者的多模态情绪数据库(multimodal emotion database of post-stroke speech dysfunction patients in Xugou, Shanxi, jointly collected by Taiyuan University of Technology and Taiyuan Central Hospital, THE-POSSD)[15],以及由伊利诺伊大学录制的构音障碍患者开源语音数据库(universal access database,UASpeech)[16]。
对中文构音障碍患者的研究正处于起步阶段,中文构音障碍数据集的缺乏一定程度上限制了国内相关研究的发展[17]。THE-POSSD数据集共采集68名被试者语音样本,68名被试者中有29名健康对照人群和39名构音障碍患者,无任何发音和听力损伤及病史,39名构音障碍患者由言语治疗康复科的临床专家根据构音障碍评定表进行评估划分为轻度、中度、重度。语料内容依据北京大学第一医院太原中心医院康复科言语治疗经验,分为非词、字、词、限制句,每人共产生65条录制音频,剔除无效语音后共计4 290条,其中对照组1 839条,轻度患者1 006条,中度患者633条,重度患者812条。在采集过程中保证安静、无电磁等外部因素干扰,且受试者均被要求完成《症状自评量表SCL-90》调查、简易精神状态检查量表(mini-mental state examination,MMSE)和Frenchay构音障碍评定表测评,并签署知情同意书,具体信息如表1所示。
National Health Commission of the People’s Republic of China. A summary of the China Stroke Prevention and Treatment Report 2021[J]. Chinese Journal of Cerebrovascular Diseases, 2023, 20(11): 783-793.
[4]
SaskaT, SudarsanaR K, PaavoA. The effect of the MFCC frame length in automatic voice pathology detection[J]. Journal of Voice: Official Journal of the Voice Foundation, 2022, 38(5): 975-982.
LiangZhengyou, LiYuxing, SunYu, et al. Speech recognition of dysarthria based on multi feature combination[J]. Computer Engineering and Design, 2022, 43(2): 567-572.
[7]
Fernández-DíazM, Gallardo-AntolínA. An attention long short-term memory based system for automatic classification of speech intelligibility[J]. Engineering Applications of Artificial Intelligence, 2020, 96.DOI: 10.1016/j.engappai.2020.103976.
[8]
YeW J, JiangZ X, LiQ, et al. A hybrid model for pathological voice recognition of post-stroke dysarthria by using 1DCNN and double-LSTM networks[J]. Applied Acoustics, 2022, 197.DOI: 10.1016/j.apacoust.2022.108934.
[9]
JoshyA A, RajanR. Dysarthria severity classification using multi-head attention and multi-task learning[J]. Speech Communication, 2023, 147: 1-11.
XueYajie, HeHongxia, YangYi. Speech signal recognition and classification based on neural network[J]. Modern Electronics Technique, 2023, 46(24): 79-84.
[14]
HeL, FuJ, LiY, et al. WNSA-net: an axial-attention-based network for schizophrenia detection using wideband and narrowbandspectrograms[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2022, 31: 721-733.
LiaoGuangkai, Zhangzheng, SongZhiguo. Convolutional network-based vehicle re-identification combining wavelet features and attention mechanism[J]. Journal of Computer Applications, 2022, 42(6): 1876-1883.
[18]
LiJ, WenY, HeL. Scconv: spatial and channel reconstruction convolution for feature redundancy[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2023: 6153-6162.
[19]
OuyangD, HeS, ZhangG, et al. Efficient multi-scale attention module with cross-spatial learning[C]// IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP). New York: IEEE, 2023: 1-5.
[20]
ZhuT, DuanS, LiangH, et al. Design, construction and evaluation of emotional multimodal pathological speech database[J]. arXiv Preprint arXiv: 2312.08998, 2023.
[21]
ChowdaryP N, AkshayM S, AravindV S, et al. A few-shot approach to dysarthric speech intelligibility level classification using transformers[C]// 14th International Conference on Computing Communication and Networking Technologies(ICCCNT) New York: IEEE, 2023: 1-6.
SongWei, ZhangYanghao. A survey of specific speech recognition algorithms for dysarthria[J]. Computer Engineering and Applications, 2024, 60(11): 62-74.
YuanLu, TaoQing, LiuJingxuan, et al. Improved dual-stream multi-modal information fusion method for sitting posture recognition[J]. Science Technology and Engineering, 2024, 24(5): 1980-1988.
[26]
YaoL X, LuY, QianY K, et al. High-accuracy classification of multiple distinct human emotions using EEG differential entropy features and ResNet18[J]. Applied Sciences, 2024, 14(14). DOI: 10.3390/app14146175.
[27]
NarendraN P, AlkuP. Glottal source information for pathological voice detection[J]. IEEE Access, 2020, 8: 67745-67755.
[28]
JavanmardiF, TirronenS, KodaliM, et al. Wav2vec-based detection and severity level classification of dysarthria from speech[C]// IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP). New York: IEEE, 2023: 1-5.