A Study of Speech Recognition Techniques for Dysarthria Speeches Based on Digit Recognition

With the introduction of speech recognition technology into people's lives, the application of which is no longer limited to ordinary people, but begins to target special populations, such as patients with dysarthria. It has become a major research challenge to build a high-performance speech recognition system for patients with dysarthria. To address this problem, in this paper the speech data of four patients with dysarthria and four healthy speakers on digital commands are firstly collected, secondly several speech recognition experiments are separately conducted by using Aishell-2 and Wenetspeech, two pre-trained models based on the WeNet open-source architecture, as well as the four major commercial recognition API systems, at the end a method to optimize the dysarthria data by using the so-vits-svc tool is proposed. The experiments show that the current state-of-the-art speech recognition model has a high recognition rate for speech data from the general population, but there is still room for improvement in the recognition performance of speech data from individuals with dysarthria. Later the optimization method for the dysarthria data proposed in this paper effectively improves the quality of speech data. For the two above open-source pre-trained models, the recognition rates have increased by 60% and 67.64 % compared to the original data, respectively.