Gradually Changing Gender Attribution of Speech Recordings Using Interpolated Filters

Abstract: When we listen to human speech, one of the first characteristics we assess is the gender of the speaker. For individuals who suffer from gender dysphoria, this may cause them to be negatively impacted by their voice not matching their gender identity. Therefore, some persons attempt to change their voices with a speech-language therapist. Differences between the average female and male voice have been studied extensively, and the findings are used in therapy to appropriately modify patients’ voices. By applying this knowledge to digitally alter patients’ voice recordings to sound more like their respective target voices, treatment could be made easier and more effective. This thesis explores the use of interpolated all-pole filters and TD-PSOLA to transform voice recordings of the vowel /a/ to be perceived as more feminine or masculine, while simultaneously attempting to preserve the qualities that make voices sound natural. Additionally, methods of measuring the distance between speech signals using the 2-Wasserstein metric are investigated. An online survey is conducted to evaluate the perceived gender and naturalness of 15 transformations. Results from the survey indicate that the gender attribution of the recordings changes when they are transformed and that the average gender scores correlate with transformation goals. It is found that five out of eleven transformed speech signals were rated as natural by more than 50 % of listeners. Furthermore, the ratings imply that several of the transformed signals were as natural sounding as unmodified ones. In conclusion, this method of voice transformation shows promise, but additional research is required before real-world applications can be made.