Implementation:Facebookresearch Audiocraft NoiseSchedule

Knowledge Sources	Facebookresearch_Audiocraft DDPM
Domains	Diffusion, Audio_Generation
Last Updated	2026-02-14 01:00 GMT

Overview

Concrete tool for implementing the noise schedule, forward diffusion process, and reverse denoising sampling for DDPM-based audio diffusion models.

Description

NoiseSchedule implements the full diffusion pipeline including forward noising (adding noise according to a beta schedule), training item generation (creating noisy input/target pairs), and reverse sampling (iterative denoising). It supports both linear and cosine beta schedules, optional noise floor, and sub-sampled step generation for faster inference.

Usage

Import this class when building or training diffusion-based audio generation models within the Multi-Band Diffusion framework.

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/modules/diffusion_schedule.py
• Lines: 1-272

Signature

class NoiseSchedule(nn.Module):
    def __init__(self, beta_t0=1e-4, beta_t1=0.02, num_steps=1000, variance="beta",
                 clip=20, rescale=1, device='cuda', beta_schedule='linear',
                 n_bands=None, sample_processor=None, noise_floor=0, **kwargs): ...

    def get_training_item(self, x: torch.Tensor, tensor_step: bool = False): ...
    def generate(self, model, initial, condition=None, return_list=False): ...
    def generate_subsampled(self, model, initial, step_list=None, condition=None, ...): ...

Import

from audiocraft.modules.diffusion_schedule import NoiseSchedule

I/O Contract

Inputs

Name	Type	Required	Description
x	torch.Tensor	Yes	Clean audio for forward noising [B, C, T]
model	nn.Module	Yes	Score/denoising model (for generation)
initial	torch.Tensor	Yes	Initial noise for reverse process [B, C, T]
condition	torch.Tensor	No	Conditioning signal

Outputs

Name	Type	Description
training_item	tuple	(noisy_input, target, step) for training
generated	torch.Tensor	Denoised audio [B, C, T]

Related Pages

• Principle:Facebookresearch_Audiocraft_Diffusion_UNet_Architecture