Implementation:Facebookresearch Audiocraft StreamingModule

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

Overview

Concrete tool for enabling incremental/causal processing of audio sequences where modules maintain internal state between calls.

Description

StreamingModule and StreamingSequential provide the base infrastructure for streaming (causal) processing in AudioCraft. StreamingModule adds a state dictionary that persists between forward calls, enabling modules to process audio incrementally. StreamingSequential chains multiple streaming modules.

Usage

Import these base classes when building streaming-capable audio processing modules (convolutions, transformers, encoders).

Code Reference

Source Location

• Repository: Facebookresearch_Audiocraft
• File: audiocraft/modules/streaming.py
• Lines: 1-131

Signature

class StreamingModule(nn.Module):
    def __init__(self): ...
    def _streaming_state(self): ...
    def set_streaming_enabled(self, enabled: bool): ...
    def flush(self, x: tp.Optional[torch.Tensor] = None): ...

class StreamingSequential(StreamingModule, nn.Sequential): ...

Import

from audiocraft.modules.streaming import StreamingModule, StreamingSequential

I/O Contract

Inputs

Name	Type	Required	Description
x	torch.Tensor	Yes	Input tensor (streamed chunk)

Outputs

Name	Type	Description
output	torch.Tensor	Processed output (same or reduced length)

Related Pages

• Principle:Facebookresearch_Audiocraft_Encoder_Decoder_Architecture