Manipulating Voice Attributes by Adversarial Learning of Structured Disentangled Representations.

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Neural voice conversion (VC) has advanced significantly, enabling realistic voice identity manipulation with minimal data.
• Existing methods primarily focus on altering voice identity, leaving manipulation of other voice attributes less explored.

Purpose of the Study:

• To present an original neural architecture for voice conversion that enables manipulation of voice attributes such as gender and age.
• To disentangle speech signal information into interpretable voice attributes for flexible manipulation.

Main Methods:

• The proposed architecture is inspired by fader networks, utilizing adversarial loss to achieve mutual independence of encoded information.
• Speech signal information is disentangled into voice attributes, allowing for manipulation during inference.
• The method was evaluated on voice gender conversion using the VCTK dataset.

Main Results:

• Quantitative analysis confirmed the architecture learns gender-independent speaker representations.
• Speaker identity recognition remained accurate from the gender-independent representation.
• Subjective experiments demonstrated high efficiency and naturalness in voice gender conversion.

Conclusions:

• The novel neural architecture successfully disentangles and manipulates voice attributes, extending beyond identity conversion.
• The method achieves effective and natural-sounding voice gender conversion.
• This work opens new possibilities for nuanced voice manipulation in speech processing.