This webpage is best viewed either on Chrome or Firefox with at least 1200px resolution.For smaller screens please use Firefox only.
 
 
 
 
This is a companion webpage for the paper submitted at ICASSP 2023 titled 'Towards Controllable Audio Texture Morphing'
 
 
 
The code for this paper can be found at :
 
 
Additional details on our experiments can be found at - Supplementary Material.
 
 
 
 
 
 
 

Inter-Class Morphing (Section 4.3 in the paper)

 
Inter-class Morphing involves interpolating between two points in the class conditioning dimension of the algorithms. The samples below show inter-class morphs generated between the classes of Wind and Water using MorphGAN, One-Hot GAN and Morph2. Additionally we show some samples generated from simple waveform mixing to generate such morphs for comparison.

In the samples below, the interpolation starts at the Water class (α=0.0) and ends at Wind (α=1.0)
 
 

Morph GAN

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Sample 1











Sample 2











Sample 3











 

One Hot GAN

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Sample 1











Sample 2











Sample 3











 

Morph2

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Sample 1











Sample 2











Sample 3











 

Mix

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Sample 1











Sample 2











Sample 3











 
 
 
 
 
 
 

Intra-Class Morphing (Section 4.2 in the paper)

 
Intra-class Morphing involves interpolating between two points in the control parameters conditions of the algorithms. The samples below show intra-class morphs generated by interpolating between the lowest and the highest wind-strengths and fill-levels for Wind and Water respectively using MorphGAN, One-Hot GAN and Morph2. Additionally we show some samples generated from simple waveform mixing to generate such morphs for comparison.

In the samples below, the interpolation starts at the fill-level or wind-strength 0.0 (lowest) and ends at 1.0 (highest).
 
 

Morph GAN

fill-level/strength = 0.0
fill-level/strength = 0.1
fill-level/strength = 0.2
fill-level/strength = 0.3
fill-level/strength = 0.4
fill-level/strength = 0.5
fill-level/strength = 0.6
fill-level/strength = 0.7
fill-level/strength = 0.8
fill-level/strength = 0.9
fill-level/strength = 1.0
Sample 2











Sample 1











 

One Hot GAN

fill-level/strength = 0.0
fill-level/strength = 0.1
fill-level/strength = 0.2
fill-level/strength = 0.3
fill-level/strength = 0.4
fill-level/strength = 0.5
fill-level/strength = 0.6
fill-level/strength = 0.7
fill-level/strength = 0.8
fill-level/strength = 0.9
fill-level/strength = 1.0
Sample 1











Sample 2











 

Morph2

fill-level/strength = 0.0
fill-level/strength = 0.1
fill-level/strength = 0.2
fill-level/strength = 0.3
fill-level/strength = 0.4
fill-level/strength = 0.5
fill-level/strength = 0.6
fill-level/strength = 0.7
fill-level/strength = 0.8
fill-level/strength = 0.9
fill-level/strength = 1.0
Sample 1











Sample 2











 

Mix

fill-level/strength = 0.0
fill-level/strength = 0.1
fill-level/strength = 0.2
fill-level/strength = 0.3
fill-level/strength = 0.4
fill-level/strength = 0.5
fill-level/strength = 0.6
fill-level/strength = 0.7
fill-level/strength = 0.8
fill-level/strength = 0.9
fill-level/strength = 1.0
Sample 1











Sample 2











 
 
 
 
 
 

Audio Quality (Section 4.1 in the paper)

 
Water Wind
Training Data Samples
Original One-Hot [15, 17]
One-Hot GAN
Morph GAN
 
 
 
 
 
 
 

Exploring Semantic Information of Inter-Class conditional parameters (Section 4.4 in the paper)

 
To analyse the semantic control of the three class parameter dimensions C of MorphGAN, we varied each dimension from 0 to 1 at steps of 0.1, while keeping all other dimensions constant at 0.5, and fixing a random Z vector. Through audition and spectrogram viewing, we can describe variation in the first C dimension 0 as taking the texture from a gurgly-wind sound to a wind-like sound. Lower parameter values also contain higher frequency components from water sounds. Dimension 1 variation makes the texture go from a windy whooshing sound to a more watery swish-like sound, where the higher values of this dimension introduce the higher frequency components but at lower amplitudes. And dimension 2 variation changes the texture from water-like to wind-like.
 
 

Example 1

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Dimension 0











Dimension 1











Dimension 2











 

Example 2

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Dimension 0











Dimension 1











Dimension 2











 

Example 3

α=0.0
α=0.1
α=0.2
α=0.3
α=0.4
α=0.5
α=0.6
α=0.7
α=0.8
α=0.9
α=1.0
Dimension 0











Dimension 1











Dimension 2