Principle:Facebookresearch Audiocraft Loudness Ratio Loss

Knowledge Sources	Facebookresearch_Audiocraft
Domains	Audio_Processing, Loss_Functions
Last Updated	2026-02-14 01:00 GMT

Overview

A perceptual loss function family that measures the noise-to-signal loudness ratio across time, frequency, or time-frequency cells with softmax-weighted aggregation.

Description

Loudness Ratio Losses compute the perceptual SNR between output and reference audio by measuring loudness in time frames, frequency bands, or both. A softmax weighting scheme ensures that the noisiest regions contribute most to the loss, focusing optimization on the most perceptually problematic areas.

Usage

Use these losses in audio compression or watermarking training where perceptual quality matters more than simple MSE reconstruction.

Theoretical Basis

The loss computes noise loudness L_n and reference loudness L_r in each cell, then aggregates with softmax weighting:

${ℒ}=\sum _i\text{softmax}(L_{n,i})\cdot (L_{n,i}-L_{r,i})$

Related Pages

• Implementation:Facebookresearch_Audiocraft_LoudnessLoss