Figure 1 shows the input image and its attribution score obtained using three XAI methods for image classification based on VGG16 trained upon ImageNet data. However, current XAI methods are not perfect, and often fail to separate an object from the background, eventually only highlighting the high-contrast object contour. If the classification accuracy is high enough, the ensuing XAI result also tends to proceed properly. I also tried adapting an approach mentioned in Transfer Learning from Speaker Verification to Multispeaker Text-To-Speech Synthesis however, again, the results were not too great (though I think this was due to implementation errors made on my part and just insufficient amounts of data).As mentioned earlier, visualized interpretation with XAI methods for image classification that provides the output in the form of a heatmap is often acceptable. The result: well., you could certainly hear elements of the target voice but it wasn't immediately recognizable. One approach to this, which actually seemed to have 'reasonable' results, was to randomly initialize a CNN with random parameters and train it to output the target Mel spectogram. My first though was to train a model solely responsible for converting this 'generic' voice to the target voice (i.e. The idea to apply transfer learning didn't actually come until I experimented with a few different approaches to voice style conversion. And then simply use this model as a starting point. Of course, that's only if you want to train it from scratch however, I can leverage the REALLY BIG open-source speech datasets and train a generic model to be very good. Still, even using this model, I'd need A LOT of data to clone someone's voice to a decent quality. After trying a lot of different models (Tacotron 1/2, WaveNet, DeepSpeech), I settled upon this rather nifty approach using convolution neural networks-which in and of itself trains a lot faster than its recurrent counterparts. To be honest, I still have no idea what I meant by 'short'-perhaps 5-10 seconds? Yea, it doesn't take a scientist to figure out that that IS NOT easy. When I first started this project, my goal, while lofty ( ^ _ ^ ), was to be able to replicate someone's voice given a short audio clip. I made a video about the project and it features a Hamlet Soliloquy (my voice) and Obama reading the Navy Seal Copypasta.

Deepfake text to speech free