A Cross-Subject sEMG-to-Speech Conversion System Using Content Features and Model Calibration

Current sEMG-based speech generation methods primarily rely on large-scale datasets from single participants, which imposes a burden on users. Moreover, previous research methods often require synchronous recordings of sEMG speech for model training, making them unsuitable for patients with speech impairments. This article introduces a cross subject sEMG-to-speech (ETS) conversion system based on content features and model calibration methods. This method uses a pre trained acoustic model to extract speaker independent acoustic features, and adapts the model to the EMG features of new individuals through Child Tune model calibration. In order to develop an ETS system suitable for people with language barriers, we recommend using electronic synthesized audio to train the ETS model as a substitute for human voice, and reconstructing speech through a voice encoder with speech conversion function. The experimental results show that our proposed sEMG-to-speech (ETS) conversion system can achieve a CER of 21.71% after calibration with 20 minutes of data from new users, and using electronic synthesized audio to train the ETS model can achieve a CER equivalent to using human voice.