Omni-Embed-Audio (OEA) — Leveraging Multimodal LLMs for Robust Audio-Text Retrieval

Official code, UIQ benchmark, and pretrained checkpoints for the ACL 2026 (Oral) paper Omni-Embed-Audio: Leveraging Multimodal LLMs for Robust Audio-Text Retrieval by HaeJun Yoo, Yongseop Shin, Insung Lee, Myoung-Wan Koo, and Du-Seong Chang (Sogang University).

• 🤗 Pretrained checkpoints: https://huggingface.co/JudeJiwoo
• 🌐 Interactive web demo: https://omni-embed-audio.github.io
• 📦 UIQ benchmark (this repo): data/UIQ/
• 💻 Minimal usage example: examples/encode_example.py · examples/encode_example.ipynb

---

TL;DR

OEA is a retrieval-oriented encoder built on a multimodal LLM backbone (Qwen2.5-Omni-3B/7B, Nemotron-3B) with a single shared transformer for text and audio, plus LoRA adapters and 512-d projection heads. We additionally release User-Intent Queries (UIQ) — five query formulations (question, imperative, tagging, paraphrase, negative) over AudioCaps / Clotho / MECAT — to evaluate retrieval robustness beyond caption-style queries, and a hard-negative discrimination metric (HNSR / TFR).

Key results from the paper:

• Competitive text-to-audio retrieval vs. M2D-CLAP.
• Dominant text-to-text retrieval (+5.5%p mean R@5).
• Substantially better hard-negative discrimination (+4.3%p HNSR@10, +34.7% relative TFR@10).

---

Repository layout

AudioRetrieval/         # Python package — see AudioRetrieval/__init__.py
  cli/                  # `python -m AudioRetrieval ...` entry point
  conf/                 # Hydra configs (model / dataset / training / reranker / uiq)
  evaluation/           # Retrieval evaluation with R@k, MAP, HNSR, TFR, Δ-Rank
  embeddings/           # Embedding cache utilities
  models/               # Model adapters: OEA, LAION-CLAP, MGA-CLAP, M2D-CLAP, WavCaps, …
  preprocessing/        # Embedding precomputation + hard-negative mining
  rerankers/            # Second-stage rerankers (XACLE, late-interaction, …)
  training/oea/         # OEA LoRA + projection-head training
  uiq_generation/       # UIQ generation (GPT-backed)
  uiq_toolkit/          # Lightweight standalone UIQ generation toolkit
data/UIQ/               # Released UIQ benchmark (5 query types × 3 datasets)
examples/               # Minimal encoding example (.py + .ipynb) + 5 Clotho clips
requirements.txt

Installation

git clone https://github.com/JudeJiwoo/AudioRetrieval.git
cd AudioRetrieval
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

For CUDA 12.4: pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu124.

Quick start: encode audio & text with a pretrained OEA model

The bundled example downloads a checkpoint from Hugging Face, attaches LoRA + projection heads, and encodes the five Clotho clips under examples/data/clotho_samples/ along with five UIQ-style queries.

python examples/encode_example.py                          # default: OEA-Qwen3B-Cl
python examples/encode_example.py --model OEA-Qwen7B-Cl    # 7B variant
python examples/encode_example.py --device cpu             # CPU fallback (slow)

Or open the notebook: examples/encode_example.ipynb.

Released checkpoints (https://huggingface.co/JudeJiwoo)

Checkpoint	Base model	Trained on
OEA-Qwen3B-Cl	Qwen/Qwen2.5-Omni-3B	Clotho
OEA-Qwen3B-AC	Qwen/Qwen2.5-Omni-3B	AudioCaps
OEA-Qwen7B-Cl	Qwen/Qwen2.5-Omni-7B	Clotho
OEA-Qwen7B-AC	Qwen/Qwen2.5-Omni-7B	AudioCaps
OEA-Nemo3B-Cl	nvidia/omni-embed-nemotron-3b	Clotho
OEA-Nemo3B-AC	nvidia/omni-embed-nemotron-3b	AudioCaps

Each repo contains step_40.pt, a dict with lora_state_dict, audio_head, text_head (and training metadata). The base model is downloaded separately on first use.

UIQ benchmark

The release contains 13,053 queries — five formulations per audio clip across three datasets:

Dataset	question	imperative	paraphrase	tagging	negative
AudioCaps (test)	975	975	975	975	630
Clotho (evaluation)	1 045	1 045	1 045	1 045	542
MECAT	848	848	848	848	409

Files live under data/UIQ/ as JSONL with the schema:

{
  "audio_id":         "Santa Motor.wav",
  "dataset":          "clotho",
  "dataset_slug":     "clotho_evaluation",
  "query_type":       "question",
  "generated_query":  "Can you find audio of a machine whining and squealing while stamping?",
  "original_captions": ["...", "..."],
  "metadata":         {"split": "evaluation", "num_captions": 5},
  "source_model":     "gpt-5.1",
  "regen_model":      "gpt-5.1"
}

Negative queries additionally reference a paired hard-negative clip; see data/UIQ/README.md.

To regenerate UIQs from scratch, see AudioRetrieval/uiq_toolkit/ or AudioRetrieval/uiq_generation/.

Evaluating a model

# Baseline retrieval (text → audio, audio → text, text → text)
python -m AudioRetrieval evaluate model=omni_embed dataset=clotho

# UIQ evaluation with all five query types
python -m AudioRetrieval evaluate model=omni_embed dataset=clotho \
    eval.tasks='[text_to_audio]' uiq=clotho_all

Hydra configs under AudioRetrieval/conf/ cover the supported models (laion_clap, mga_clap, cacophony, wavcaps, omni_embed, omni_embed_7b) and datasets (clotho, audiocaps).

Training your own OEA model

See AudioRetrieval/training/oea/ — train_omniembed_lora.py is the entry point and run_training.sh shows a typical multi-stage launch. Training is parameter-efficient: only LoRA adapters (~11–16M params) and projection heads are updated; backbone weights stay frozen.

Citation

@inproceedings{yoo2026omniembedaudio,
  title     = {Omni-Embed-Audio: Leveraging Multimodal LLMs for Robust Audio-Text Retrieval},
  author    = {Yoo, HaeJun and Shin, Yongseop and Lee, Insung and Koo, Myoung-Wan and Chang, Du-Seong},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL)},
  note      = {Oral presentation},
  year      = {2026}
}

Acknowledgments

This work was partly supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2025-25441313, Professional AI Talent Development Program for Multimodal AI Agents, Contribution: 50%). This research was also supported by the MSIT, Korea, under the Top-Tier AI Global HRD invitation program (RS-2025-25461932) supervised by the IITP.

License

• Code (this repository) — MIT.
• UIQ queries under data/UIQ/ — released under CC BY 4.0. The underlying audio (AudioCaps / Clotho / MECAT) is not redistributed here; obtain it from the original sources under their respective licenses.
• Bundled Clotho samples under examples/data/clotho_samples/ — each clip retains its original Freesound license (CC0 1.0 or CC-BY 3.0). See examples/data/clotho_samples/captions.jsonl for per-clip attribution.

Acknowledgements

• Clotho — Drossos, Lipping, Virtanen (2020). Clotho: An audio captioning dataset. ICASSP 2020. https://zenodo.org/records/4783391
• AudioCaps — Kim et al. (2019). AudioCaps: Generating captions for audios in the wild. NAACL 2019.
• MECAT — multilingual & multi-event captioning benchmark.
• Base model weights are © their respective authors: Qwen2.5-Omni (Alibaba) and Omni-Embed-Nemotron-3B (NVIDIA). Please respect their licenses when redistributing derivatives.