Implementation:Facebookresearch Audiocraft LoudnessLoss

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

Overview

Concrete tool for computing perceptual loudness-ratio-based losses between output and reference audio signals in time, frequency, and combined time-frequency domains.

Description

This module implements three loudness-based loss functions: FLoudnessRatio (frequency-domain SNR across mel bands), TLoudnessRatio (time-domain SNR across overlapping frames), and TFLoudnessRatio (combined time-frequency loudness ratio). All losses use softmax weighting so that the noisiest regions dominate the loss, providing perceptually-focused training signals for audio compression and watermarking.

Usage

Import these loss functions when training audio compression or watermarking models where perceptual quality assessment based on loudness ratios is needed.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/losses/loudnessloss.py
• Lines: 1-204

Signature

class FLoudnessRatio(nn.Module):
    def __init__(self, sample_rate=16000, segment=20, overlap=0.5, epsilon=..., n_bands=0):
        """Frequency-domain loudness ratio loss."""
    def forward(self, out_sig: torch.Tensor, ref_sig: torch.Tensor) -> torch.Tensor: ...

class TLoudnessRatio(nn.Module):
    def __init__(self, sample_rate=16000, segment=0.5, overlap=0.5):
        """Time-domain loudness ratio loss."""
    def forward(self, out_sig: torch.Tensor, ref_sig: torch.Tensor) -> torch.Tensor: ...

class TFLoudnessRatio(nn.Module):
    def __init__(self, sample_rate=16000, segment=0.5, overlap=0.5, n_bands=0, ...):
        """Combined time-frequency loudness ratio loss."""
    def forward(self, out_sig: torch.Tensor, ref_sig: torch.Tensor) -> torch.Tensor: ...

Import

from audiocraft.losses.loudnessloss import FLoudnessRatio, TLoudnessRatio, TFLoudnessRatio

I/O Contract

Inputs

Name	Type	Required	Description
out_sig	torch.Tensor	Yes	Output audio signal [B, C, T]
ref_sig	torch.Tensor	Yes	Reference audio signal [B, C, T]

Outputs

Name	Type	Description
loss	torch.Tensor	Scalar loudness ratio loss

Usage Examples

from audiocraft.losses.loudnessloss import TFLoudnessRatio
import torch

loss_fn = TFLoudnessRatio(sample_rate=16000, n_bands=32)
output = torch.randn(4, 1, 16000)
reference = torch.randn(4, 1, 16000)
loss = loss_fn(output, reference)

Related Pages

• Principle:Facebookresearch_Audiocraft_Loudness_Ratio_Loss