Abstract

In this project, a machine critic is built to give comments on several pictures. The critic is not intelligent enough to surpass describing the contents of pictures. However, it performs excellent tongue-twisters to beat human beings. The voice of the critic comes from two models, DeepVoice3 and Tacotron2, respectively.

Results

106 - DeepVoice3

106 - Tacotron2

8 - DeepVoice3

8 - Tacotron2

101 - DeepVoice3

101 - Tacotron2

Tongue-twister 1: Peter Piper picked a peck of pickled peppers. How many pickled peppers did Peter Piper pick?

DeepVoice3

Tacotron2

Tongue-twister 2: A skunk sat on a stump and thunk the stump stunk, but the stump thunk the skunk stunk.
DeepVoice3

Tacotron2

Code

• DeepVoice3.ipynb – The first voice
• Tacotron2_and_WaveNet.ipynb – The second voice

Technical Notes

• All the packages are included when running ipynb files.
• Tested on Google Colab.

Reference

• DeepVoice3 implementation
• DeepVoice3 Audio samples, for details, click here
• DeepVoice3 Google Colab example
• Image captioning
• Tacotron2 (mel-spectrogram prediction part)
• WaveNet