Principle:Facebookresearch Audiocraft Multi Period Discrimination
| Knowledge Sources | |
|---|---|
| Domains | Audio_Synthesis, GAN |
| Last Updated | 2026-02-14 01:00 GMT |
Overview
An adversarial discrimination technique that reshapes audio into 2D periodic representations to capture harmonic patterns at different fundamental periods.
Description
Multi-Period Discrimination analyzes audio quality by exploiting the periodic nature of audio signals. The 1D waveform is reshaped into 2D tensors using different period values (corresponding to different fundamental frequencies), and 2D convolutions assess the periodic structure. Using coprime periods (e.g., [2,3,5,7,11]) ensures coverage of diverse harmonic patterns without redundancy.
Usage
Use this principle in adversarial training for audio synthesis or compression when you need a discriminator that is sensitive to harmonic artifacts and periodic distortions.
Theoretical Basis
Given a 1D signal x of length T and period p, reshape into a 2D tensor of shape (T/p, p), then apply 2D convolutions. Each period captures a different harmonic structure.