AMBIT model checkpoint

Code and instructions · Demo · Paper

AMBIT supports text-guided First-Order Ambisonics (FOA) generation and editing. Editing additionally takes reference FOA audio.

One unified model supports both generation and editing. AR and DiT share the same Transformer backbone. All necessary model assets are included in the download.

Use

Requires Python 3.10+, Linux, and a CUDA GPU with bfloat16 support. Install the matching inference code, then download this repository:

git clone https://github.com/WJM-George/AMBIT.git
cd AMBIT
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
hf download WJM-George/AMBIT --local-dir models/AMBIT

Generation:

python -m ambit.infer generate --bundle models/AMBIT \
  --prompt "A dog barks to the left while a car passes from front to right." \
  --output outputs/generated.wav --seed 42

Editing:

python -m ambit.infer edit --bundle models/AMBIT \
  --source input_foa.wav \
  --prompt "Move the dog to the right while keeping the other sounds unchanged." \
  --output outputs/edited.wav --seed 42

Both commands save FOA audio and the generated ScenePlan. Editing uses a 44.1 kHz, four-channel ACN/SN3D FOA reference in W, Y, Z, X order.

Files

• model.safetensors: shared Transformer, AR head, conditioning and AMBIT CLAP weights.
• vae.safetensors: FOA audio encoder/decoder.
• bundle.json, renderer.json, vae.json: inference configurations.
• codec/: ScenePlan tokenization assets.
• planner_encoder/, text_encoder/: text encoders and tokenizers.

License and attribution

AMBIT is distributed under the MIT license. Included third-party components retain their respective licenses; see LICENSES/. AMBIT builds on stable-audio-tools.