Audio2Guitarist-GAN

A two-stage generative adversarial network that generates images of guitarists playing guitar from audio.

Descriptions

To be updated.

Architecture

Stage 1: Audio to binary mask  

Stage 2: Binary mask to color image

More information in this blog post.

Result

1. Video output

• Case 1: Test on the same guitarist, 南澤大介.
  • Video 1: トゥ・ビー・ウィズ・ユー (acoustic guitar solo)
    • Source YouTube video here.
  • Video 2: 愛はかげろうのように (acoustic guitar solo)
    • Source YouTube video here.
• Case 2: Test on different guitarist, 伍々慧. The playing style and recording are different from the training data.
  • Video 3: Autumn Leaves (early version) / Satoshi Gogo
    • Source YouTube video here.
  • Video 4: I got rhythm / Satoshi Gogo
    • Source YouTube video here.
• Case 3: Test on different instruments and human voice.
  • [TBU]

Here are the official website of 南澤大介 and 伍々慧.

2. Conditional output

The following gifs are result images generated from an audio that the model had never seen.

• Source video (audio): tupliのテーマ (acoustic guitar solo) 作曲／編曲：南澤大介
• Top to bottom: audio visualization, stage-1 output, stage-2 output, ground truth.

3. Pose-guided generation

The following gifs show outputs of 2nd-stage model given conditional poses.

• Source video (audio): John Pizzarelli - "I Got Rhythm" (solo) at the Fretboard Journal
• Top: Reference video; Middle: conditional hands input; Bottom: stage 2 output.

References

1. Audio to Body Dynamics
2. Pose guided person image generation
3. Deep Video Generation, Prediction and Completion of Human Action Sequences
4. Deformable GANs for Pose-based Human Image Generation
5. Skeleton-aided Articulated Motion Generation
6. Assessment of Student Music Performances Using Deep Neural Networks
7. Dance Dance Convolution