The invention discloses a
stroke dysarthria speech recognition method and device, and the technical scheme of the invention comprises the steps: carrying out the classification of obtained speech sample data according to the
syllable type represented by audio, normal and patient, obtaining an audio
spectrogram through transformation, and carrying out the construction of a
network model, thereby achieving the recognition of a
stroke dysarthria speech. The front-end network utilizes
plant morphological
physiology to design a core
processing module, continuous Downsample behind the STEM module forms a
rhizome so as to quickly transport calculation nodes to a higher
receptive field, and the rear-end network forms a
vine cross
structure based on alternate configuration of depth separable
convolution and attention mechanisms based on the Xception module so as to quickly transport the calculation nodes to a higher
receptive field. An attention mechanism is selectively arranged in an alternating
convolution module, so that the recognition capability and accuracy of key speech features are improved, a global
receptive field can be captured on multiple scales, and cerebral apoplexy
pathologic significant feature information can be accurately learned and discriminated; the technical problem of low accuracy of the existing
stroke patient pronunciation analysis is solved.