transaudio 0.1.3

Transferable Audio Representations via Procedural Generation (Physics-based AudioMAE)

TransAudio: Transferable Audio Representations via Procedural Generation

An implementation of the AudioMAE model pre-trained entirely on procedurally generated audio (physics-based synthesis). This library allows you to train powerful audio representations without using any real-world data, achieving strong sim-to-real transfer.

Based on the paper: Transferable Audio Representations via Procedural Generation[cite: 1].

Features

• Physics-based Synthesizer : Generates infinite training data using Additive, FM, and Pulse synthesis with ADSR envelopes and transient bursts[cite: 1].
• AudioMAE : Masked Autoencoder tailored for audio spectrograms.
• Zero-Data Pre-training : Train models without downloading massive datasets like AudioSet.

Installation

pip install transaudio 

Usage

1. Training from Scratch

You can start the pre-training loop immediately after installation. The synthesizer runs on-the-fly.

transaudio-train --epochs 100 --batch_size 32

2. Using the Model in Python

import  torch 
from transaudio import  mae_vit_base_patch16, advanced_physics_synth 
 
# 1. Create the model 
model = mae_vit_base_patch16(audio_exp=True, img_size=(1024, 128 )) 
 
# 2. Generate a synthetic sample 
waveform = advanced_physics_synth(f0_hz=440.0 ) 
print(f"Generated waveform shape: {waveform.shape}" ) 
 
# 3. Forward pass (Simulate input) 
# Input expected: [Batch, Channel, Time, Freq] 
dummy_input = torch.randn(1, 1, 1024, 128 ) 
loss, pred, mask = model(dummy_input, mask_ratio=0.75 ) 
print(f"Loss: {loss.item()}" ) 

Citation

If you use this code, please cite the original paper:

IEEE/ACM TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. XX, NO. XX, 2026 
Transferable Audio Representations via Procedural Generation 
Fengrui Liu, Ruiyang Huang, Qijian Zheng, Yuanfang Wang, and Feng Liu