Investigation of Vowel Generation Method in Low-resource Pathological Voice Database

The number of voices in commonly used pathological voice databases such as Massachusetts Eye and Ear Infirmary (MEEI) and Saarbruecken Voice Database (SVD) is insufficient and imbalanced. This may make the classification results lack credibility and robustness. Our works investigate the direct generation of normally pitched vowel /:a/ to expand the low-resource pathological voice database. A framework for generating vowels of different lengths using improved WaveNet and Generative Adversarial Network (GAN) is proposed in this work. Long and short vowel segments can be generated by an improved WaveNet model and a model based on Stationary Wavelet Transform and Wasserstein GAN with gradient-penalty (SWT-WGANGP) in our framework respectively. The generated voice segments are added to the original imbalanced database to improve the classification performance. Besides, we propose the accuracy -score and the diversity-score to evaluate the generated voice. By adding our generated data based on the traditional classification pipeline, accuracy on the two databases increased by 1.81% and 4.38% respectively, and the recall has achieved more than 10% improvement. Compared with the other data generation method, our method improves the classification results the most while using the same feature and classifier. Besides, using our proposed framework based on existing advanced pathological voice detection methods can further improve performance. Our results show that deep generative models with optimized structure can be used for direct vowel generation in low-resource pathological voice databases to expand the raw database, which has been ignored in previous research. Our framework can be used as a generic pre-processing module to improve the detection of pathological voice.