Physician-patient speech separation method based on voiceprint technology and privacy protection

Abstract

In speech recognition and natural language processing of doctor-patient voice communication, it is critical to distinguish what comes from the healthcare worker and what comes from the patient. In addition, speech contains acoustic and linguistic features that can be identified by machine learning models to measure the speaker's behavioral health. At the same time, it is relatively simple and attractive for patients to use voice data acquisition, as well as relatively cheap and convenient, requiring only a microphone, a quiet place and a device to collect audio samples. Thus, voice-based biomarkers can prescreen for disease, monitor disease progression and response to treatment, and be useful alternative markers for clinical studies with informed consent, but the premise of this process again requires us to distinguish between doctors and patients when taking audio samples. For audio samples of patient voices, in practice, most of the doctor's and patient's voices are not taken separately, but are mixed together. Several speaker recording methods have been used to isolate sound in the time domain; However, these studies do not address how to obtain timelabel-based speech samples, nor how to identify speakers. In this paper, a speech separation method is proposed for the audio separation situation between a doctor and several patients. The method mainly includes three parts: voiceprint segmentation clustering, cutting and splicing, speech identity determination. Doctor and patient audio can be separated while respecting the privacy of the conversation content, and can be stored separately based on the identity of the voice.