Parkinson’s disease patients suffer from disorders of speech.The most frequently reported speech problems are weak,hoarse,nasal or monotonous voice,imprecise articulation,slow or fast speech,difficulty starting speec...Parkinson’s disease patients suffer from disorders of speech.The most frequently reported speech problems are weak,hoarse,nasal or monotonous voice,imprecise articulation,slow or fast speech,difficulty starting speech,impaired stress or rhythm,stuttering,and tremor.To improve the speech quality and assist the patient with speech rehabilitation therapy,we have proposed the speech recognition model for Parkinson’s disease patients using transfer learning technique(PSTL),where we have pre-trained the long short-term memory(LSTM)neural network model with our developed publicly available dataset that has been obtained from healthy people through the social media platform.Then,we applied the transfer learning technique to improve the performance of the PSTL framework.The frequency spectrogram masking data augmentation method has been used to alleviate the over-fitting problem so that the word error rate(WER)is further reduced.Even with a limited dataset,our proposed model has effectively reduced the WER from 58% to 44.5% on the original speech dataset and 53.1% to 43% on the denoised speech dataset,which demonstrated the feasibility of our framework.展开更多
基金the National Key Research and Development Program of China(No.2019YFB2204500)the Science,Technology and Innovation Action Plan of Shanghai Municipality(No.1914220370)。
文摘Parkinson’s disease patients suffer from disorders of speech.The most frequently reported speech problems are weak,hoarse,nasal or monotonous voice,imprecise articulation,slow or fast speech,difficulty starting speech,impaired stress or rhythm,stuttering,and tremor.To improve the speech quality and assist the patient with speech rehabilitation therapy,we have proposed the speech recognition model for Parkinson’s disease patients using transfer learning technique(PSTL),where we have pre-trained the long short-term memory(LSTM)neural network model with our developed publicly available dataset that has been obtained from healthy people through the social media platform.Then,we applied the transfer learning technique to improve the performance of the PSTL framework.The frequency spectrogram masking data augmentation method has been used to alleviate the over-fitting problem so that the word error rate(WER)is further reduced.Even with a limited dataset,our proposed model has effectively reduced the WER from 58% to 44.5% on the original speech dataset and 53.1% to 43% on the denoised speech dataset,which demonstrated the feasibility of our framework.
