EnhancingSpeech_02456

Enhancing Speech

Can StarGANv2-VC be utilized as an intelligent audiofilter for upscaling phone conversations to be less distorted?

Proof of Concept

The StarGANv2-VC model is able to convert speech of an unrecognized input speaker. As such, we tested whether the StarGAN-v2 VC model could clean distorted audio out of the box. The utterances of one of the speakers can be heard below.

Original Audio Recording Same Clip with Distortion StarGANv2-VC,
Mapping (Distortion → Original)

As can be heard, the StarGANv2-VC model cannot remove the introduced distortion, when it has not been trained on distorted noise.

Training on Noisy Data

We added the constructed, distorted versions of the data, and trained the StarGANv2-VC model on this in addition to the original training data. We then replaced standard convolutions in the StarGANv2-VC with depthwise separable convolutions which resulted in a 63 % total reduction in the number of trainable model parameters, to conform with the aims of augmentedhearing.io to have a model better fit for mobile devices. Listen below to see if you can hear a difference! This second model version is known as the “reduced” model.

StarGANv2-VC,
Mapping (Distortion → Original)
StarGANv2-VC trained on distorted data,
Mapping (Distortion → Original)
StarGANv2-VC (reduced) trained on distorted data,
Mapping (Distortion → Original)