Audio ControlNet

This is the official repository for the paper "Audio ControlNet for Fine-Grained Audio Generation and Editing".

Audio ControlNet enables fine-grained control over audio generation through multiple conditioning mechanisms including loudness, pitch, and sound events. This allows for precise audio synthesis and editing with controllable attributes.

🤗 Huggingface Space Demo   |   🌐 Web Demo   |   🤗 HuggingFace Checkpoints

🎯 Supported Control Types

• Loudness: Controls energy dynamics and volume variations
• Pitch: Controls fundamental frequency contours
• Events: Controls timing and occurrence of specific sound categories

🚀 Quick Start

Installation

First, install the package in development mode:

pip install -e .

This will install all required dependencies including PyTorch, librosa, transformers, and other necessary packages.

Basic Usage

from audio_controlnet.infer import AudioControlNet

model = AudioControlNet.from_pretrained(MODEL_NAME)

# Generate audio with text prompt
caption = "A man is speaking while walking, music is playing, and sound effects are heard."

res = model.infer(
    caption=caption,
    control=CONTROLS,
)

📖 Usage Examples

1. Sound Events Control

Control the timing and occurrence of specific sound events:

model = AudioControlNet.from_pretrained('juhayna/T2A-Adapter-events-v1.0')

# Define sound events with timestamps
events = {
    "Sound effect": [[0.5, 1.6], [4.1, 5.4], [7.0, 8.5]], 
    "Male speech, man speaking": [[5.0, 10.0]], 
    "Music": [[9.0, 10.0]]
}

# Generate audio with events conditioning
res = model.infer(
    caption=caption,
    control={'events': events}
)
torchaudio.save('./output/events_controlled.wav', res.audio, res.sample_rate)

2. Loudness Control

Control the energy dynamics of generated audio using a reference audio file:

model = AudioControlNet.from_pretrained('juhayna/T2A-Adapter-loudness-v1.0')

res = model.infer(
    caption=caption,
    control={'loudness': model.prepare_loudness('./reference.flac')}
)
torchaudio.save('./output/loudness_controlled.wav', res.audio, res.sample_rate)

3. Pitch Control

Control the pitch contour of generated audio:

model = AudioControlNet.from_pretrained('juhayna/T2A-Adapter-pitch-v1.0')

res = model.infer(
    caption=caption,
    control={'pitch': model.prepare_pitch('./reference.flac')}
)
torchaudio.save('./output/pitch_controlled.wav', res.audio, res.sample_rate)

4. Multi-Control Generation

Combine multiple control conditions for fine-grained control:

model = AudioControlNet.from_multi_controlnets([
    'juhayna/T2A-Adapter-loudness-v1.0',
    'juhayna/T2A-Adapter-pitch-v1.0',
    'juhayna/T2A-Adapter-events-v1.0',
])

# Use multiple controls simultaneously
res = model.infer(
    caption=caption,
    control={
        'loudness': model.prepare_loudness('./reference.flac'),
        'pitch': model.prepare_pitch('./reference.flac'),
        'events': events
    }
)
torchaudio.save('./output/multi_controlled.wav', res.audio, res.sample_rate)

🎛️ Web Interface

Launch the Gradio web interface for interactive audio generation:

gradio app.py

The web interface provides:

• Text-to-audio generation
• Interactive control condition setup
• Real-time visualization of loudness, pitch, and events
• Audio playback and download

🚂 Training

Coming soon. (⏳under preparation and will be released in a few days !)

📋 Citation

@article{zhu2026audiocontrolnet,
      title={Audio ControlNet for Fine-Grained Audio Generation and Editing}, 
      author={Haina Zhu and Yao Xiao and Xiquan Li and Ziyang Ma and Jianwei Yu and Bowen Zhang and Mingqi Yang and Xie Chen},
      journal={arXiv preprint arXiv:2602.04680},
      year={2026}
}

🙏 Acknowledgements

We specially thank the following repositories:

• MeanAudio
• av-benchmark
• AudioComposer