Abstract
Laughter is a complex, non-verbal emotional signal that serves as a significant challenge for audio synthesis due to its lack of linguistic structure. This paper introduces LaughGANter, a generative adversarial network (GAN) based on the Multi-Scale Gradient (MSG-GAN) architecture, specifically designed to synthesize diverse and high-quality human laughter. By incorporating dilated convolutions for expanded receptive fields and categorical conditioning for control over age and gender, the model achieves realistic results. Evaluation using a domain-specific Fréchet Inception Distance (FID) and latent space interpolation demonstrates the model's ability to generate convincing, controllable laughter for interactive art applications.
Introduction
While artificial intelligence has made significant strides in speech synthesis, laughter remains a "final frontier." Unlike speech, which is governed by phonetic and semantic structures, laughter is a purely emotional, non-verbal signal that is highly context-dependent, varying based on age, gender, and emotional state. Most audio synthesis research focuses on speech, leaving laughter under-explored within modern GAN frameworks.
This project was motivated by both technical gaps and social needs arising from the COVID-19 crisis. The authors aimed to use "synthetic empathy" to reconnect isolated generations through an interactive art experience driven by the "welcome melody of laughter."
Methodology
The LaughGANter model is built upon the Multi-Scale Gradient GAN (MSG-GAN) architecture. Unlike standard GANs that may struggle with stability at high resolutions, MSG-GAN allows the discriminator to penalize the generator at multiple resolutions simultaneously, simplifying hyperparameter tuning compared to progressive growing techniques.
To capture the long-range correlations and rhythmic patterns inherent in laughter, the researchers implemented Induced Receptive Field Expansion. This was achieved by adding residual blocks with dilated convolutions after each upsampling layer in the generator, exponentially increasing the model's "view" of the data without a proportional increase in parameters.
To ensure controllability, the model utilizes Categorical Conditioning. Metadata such as age and gender are converted into categorical embeddings, which are concatenated with the latent noise vector in the generator and injected into the multi-scale vectors of the discriminator. The training process employed the Relativistic Average Hinge loss, encouraging the generator to produce samples that are increasingly more realistic than the current synthetic outputs.
Dataset and Evaluation
The model was trained on a specialized dataset of 2,145 laughter samples (1–8 seconds, 22.05kHz mono) provided by the National Film Board of Canada. The dataset includes a demographic mix (55% male, 45% female) and was augmented using pitch and duration shifting to prevent overfitting.
Evaluating "realistic laughter" is inherently subjective. To provide a quantitative measure, the authors developed a domain-specific FID. While standard FID relies on ImageNet (visual) classifiers, this study trained a custom classifier on gender and age specifically using laughter spectrograms to calculate a more meaningful distance metric for the audio domain.
Results and Discussion
Qualitative analysis of log-magnitude spectrograms reveals that LaughGANter successfully captures the rhythmic "bursts" characteristic of human laughter. As shown in the training progression, the model evolves from generating unstructured noise to producing clear, harmonic structures that closely mirror real samples.
Quantitatively, the custom FID score showed a consistent decrease over training epochs, indicating that the generated distribution successfully converged toward the real data. Comparison between generated and real spectrograms further validates the model's ability to synthesize realistic acoustic attributes.
Furthermore, the model demonstrated successful latent interpolation, or "laughter morphing." By transitioning between points in the latent space (e.g., between male and female identities), the model generates smooth acoustic hybrids, showcasing the power of its learned representation.
Conclusion and Ethical Considerations
While LaughGANter represents a significant step in unconditional audio synthesis, limitations remain regarding dataset size and the lack of universal evaluation metrics. The authors also highlight ethical considerations: the ability to precisely mimic emotional responses carries risks of manipulation. Consequently, this work is framed strictly for artistic purposes, emphasizing the need for ethical guardrails in emotional AI. Future work will focus on deploying these models into web-based interactive art experiences to explore the relationship between memory, empathy, and sound.
Summary updated