File size: 13,977 Bytes
c6e29fb f5101f4 96bb163 5d7aa11 c6e29fb 521c6d7 f5101f4 521c6d7 ba8f8bb 521c6d7 0a7d0a4 c6e29fb 5d7aa11 ba8f8bb c6e29fb 96bb163 c6e29fb 96bb163 c6e29fb 5d7aa11 c6e29fb 5d7aa11 c6e29fb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 |
---
license: other
language:
- en
tags:
- audio
- reasoning
- audio understanding
- ASR
- chat
- voice
arxiv: 2507.08128
datasets:
- nvidia/LongAudio
- nvidia/AudioSkills
- nvidia/AF-Think
- nvidia/AF-Chat
base_model:
- nvidia/audio-flamingo-3
---
# Model Overview
<div align="center" style="display: flex; justify-content: center; align-items: center; text-align: center;">
<a href="https://github.com/NVIDIA/audio-flamingo" style="margin-right: 20px; text-decoration: none; display: flex; align-items: center;">
<img src="static/logo-no-bg.png" alt="Audio Flamingo 3 🔥🚀🔥" width="120">
</a>
</div>
<div align="center" style="display: flex; justify-content: center; align-items: center; text-align: center;">
<h2>
Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio-Language Models
</h2>
</div>
<div align="center" style="display: flex; justify-content: center; margin-top: 10px;">
<a href="https://arxiv.org/abs/2507.08128"><img src="https://img.shields.io/badge/arXiv-2503.03983-AD1C18" style="margin-right: 5px;"></a>
<a href="https://research.nvidia.com/labs/adlr/AF3/"><img src="https://img.shields.io/badge/Demo page-228B22" style="margin-right: 5px;"></a>
<a href="https://github.com/NVIDIA/audio-flamingo"><img src='https://img.shields.io/badge/Github-Audio Flamingo 3-9C276A' style="margin-right: 5px;"></a>
<a href="https://github.com/NVIDIA/audio-flamingo/stargazers"><img src="https://img.shields.io/github/stars/NVIDIA/audio-flamingo.svg?style=social"></a>
</div>
<div align="center" style="display: flex; justify-content: center; margin-top: 10px; flex-wrap: wrap; gap: 5px;">
<a href="https://huggingface.co/nvidia/audio-flamingo-3">
<img src="https://img.shields.io/badge/🤗-Checkpoints-ED5A22.svg">
</a>
<a href="https://huggingface.co/nvidia/audio-flamingo-3-chat">
<img src="https://img.shields.io/badge/🤗-Checkpoints (Chat)-ED5A22.svg">
</a>
<a href="https://huggingface.co/datasets/nvidia/AudioSkills">
<img src="https://img.shields.io/badge/🤗-Dataset: AudioSkills--XL-ED5A22.svg">
</a>
<a href="https://huggingface.co/datasets/nvidia/LongAudio">
<img src="https://img.shields.io/badge/🤗-Dataset: LongAudio--XL-ED5A22.svg">
</a>
<a href="https://huggingface.co/datasets/nvidia/AF-Chat">
<img src="https://img.shields.io/badge/🤗-Dataset: AF--Chat-ED5A22.svg">
</a>
<a href="https://huggingface.co/datasets/nvidia/AF-Think">
<img src="https://img.shields.io/badge/🤗-Dataset: AF--Think-ED5A22.svg">
</a>
</div>
<div align="center" style="display: flex; justify-content: center; margin-top: 10px;">
<a href="https://huggingface.co/spaces/nvidia/audio-flamingo-3"><img src="https://img.shields.io/badge/🤗-Gradio Demo (7B)-5F9EA0.svg" style="margin-right: 5px;"></a>
</div>
## Description:
Audio Flamingo 3 (AF3) is a fully open, state-of-the-art Large Audio-Language Model (LALM) that advances reasoning and understanding across speech, sounds, and music. AF3 builds on previous work with innovations in:
- Unified audio representation learning (speech, sound, music)
- Flexible, on-demand chain-of-thought reasoning
- Long-context audio comprehension (up to 10 minutes)
- Multi-turn, multi-audio conversational dialogue (AF3-Chat)
- Voice-to-voice interaction (AF3-Chat)
Extensive evaluations confirm AF3’s effectiveness, setting new benchmarks on over 20 public audio understanding and reasoning tasks.
**This model is the chat version of AF3, capable of voice chat and muiti-tun multi-audio dialogue. The non-chat version can be found [here](https://huggingface.co/nvidia/audio-flamingo-3/)**
**Please note that we do not currently provide the streaming TTS-based voice output module. We plan to release it at a later date along with a detailed report.**
**This model is for non-commercial research purposes only.**
## Results:
<center><img src="static/af3_radial-1.png" width="400"></center>
## Model Architecture:
Audio Flamingo 3 uses AF-Whisper unified audio encoder, MLP-based audio adaptor, Decoder-only LLM backbone (Qwen2.5-7B), and Streaming TTS module (AF3-Chat). Audio Flamingo 3 can take up to 10 minutes of audio inputs.
<center><img src="static/af3_main_diagram-1.png" width="800"></center>
## License / Terms of Use
The model is released under the [NVIDIA OneWay Noncommercial License](static/NVIDIA_OneWay_Noncommercial_License.docx). Portions of the dataset generation are also subject to the [Qwen Research License](https://huggingface.co/Qwen/Qwen2.5-3B/blob/main/LICENSE) and OpenAI’s [Terms of Use](https://openai.com/policies/terms-of-use).
## Deployment Geography
Global.
## Use Case
Intended for researchers and developers to explore:
- Audio question answering and reasoning
- Long-context audio comprehension
- Interactive sound/music design assistants
- Multi-turn (voice) chat
## Release Date
- Github (07/10/2025) via https://github.com/NVIDIA/audio-flamingo
- HuggingFace (07/10/2025) via https://huggingface.co/nvidia/audio-flamingo-3
## References:
* [Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio-Language Models]()
* [Project Page](https://github.com/NVIDIA/audio-flamingo)
* [Demo Website](https://research.nvidia.com/labs/adlr/AF3/)
* [Hugging Face](https://huggingface.co/nvidia/audio-flamingo-3)
## Model Architecture:
**Architecture Type:** Transformer
**Network Architecture:** Audio Flamingo 3
AF3 uses:
- AF-Whisper unified audio encoder
- MLP-based audio adaptor
- Decoder-only LLM backbone (Qwen2.5-7B)
- Streaming TTS module (AF3-Chat)
**This model was developed based on [NVILA](https://github.com/NVlabs/VILA/tree/main/scripts/NVILA-Lite) and [Qwen-2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B) <br>
## Input:
- Input Type: Audio, Text <br>
- Input Format: WAV/MP3/FLAC, UTF-8 text <br>
- Input Parameters: Audio is Two-Dimensional (2D) and Text is One-Dimensional (1D)<br>
- Other Properties Related to Input: <br>
- Max Audio Length: 10 Minutes <br>
- Max Text Length: 16000 tokens<br>
## Output:
- Output Type: Text (and optional speech) <br>
- Text Format: UTF-8 string <br>
- Output Parameters: One-Dimensional (1D)<br>
- Other Properties Related to Output: <br>
- Max Text Length: 1024 tokens <br>
- Speech Format: streaming TTS (text-to-speech) waveform<br>
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems (A100/H100). By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions. <br>
## Software Integration:
**Runtime Engine:** PyTorch / HuggingFace Transformers
**Supported Hardware:**
* NVIDIA Ampere (A100)
* NVIDIA Hopper (H100)
**Supported OS:**
* Linux
## Model Version:
* v3.0
---
## Training and Testing Datasets:
### Training Dataset:
AF3 is trained entirely on open-source audio data, organized into four novel, large-scale collections. For each dataset, we mention whether the dataset annotations are collected by Human or they are Automated i.e. generated using AI models.
The data collection method noted below applies for all datasets used for training and testing:
Data Collection Method: Human
Labeling Collection Method: Please see below:
#### General Sound:
* [WavCaps](https://github.com/XinhaoMei/WavCaps) (Automated)
* [MACS](https://zenodo.org/records/5114771) (Human)
* [SoundDescs](https://github.com/akoepke/audio-retrieval-benchmark) (Human)
* [Clotho-v2](https://github.com/audio-captioning/clotho-dataset/tree/master) (Human)
* [WavText5K](https://github.com/microsoft/WavText5K) (Human)
* [Clotho-AQA](https://zenodo.org/records/6473207) (Human)
* [Open-AQA](https://github.com/YuanGongND/ltu?tab=readme-ov-file) (Automated)
* [CompA-R](https://github.com/Sreyan88/GAMA) (Automated)
* [Salmonn AQA](https://github.com/bytedance/SALMONN/tree/main) (Automated)
* [Audio Entailment](https://github.com/microsoft/AudioEntailment)(Automated)
* [CompA](https://github.com/Sreyan88/CompA) (Automated)
* [AudioSet](https://research.google.com/audioset/download.html) (Human)
* [YouTube-8M](https://research.google.com/youtube8m/) (Human)
* [FSD50k](https://zenodo.org/records/4060432) (Human)
* [CochlScene](https://github.com/cochlearai/cochlscene) (Human)
* [NonSpeech7K](https://zenodo.org/records/6967442) (Human)
* [Chime-Home](https://code.soundsoftware.ac.uk/projects/chime-home-dataset-annotation-and-baseline-evaluation-code) (Human)
* [Sonyc-UST](https://zenodo.org/records/3966543) (Human)
#### Music:
* [LP-MusicCaps](https://github.com/seungheondoh/lp-music-caps) (Automated)
* [MusicQA](https://github.com/shansongliu/MU-LLaMA?tab=readme-ov-file) (Automated)
* [MusicAVQA](https://gewu-lab.github.io/MUSIC-AVQA/) (Human)
* [MusicBench](https://huggingface.co/datasets/amaai-lab/MusicBench) (Automated)
* [Mu-LLAMA](https://github.com/shansongliu/MU-LLaMA) (Automated)
* [NSynth](https://magenta.tensorflow.org/datasets/nsynth) (Human)
* [FMA](https://github.com/mdeff/fma) (Human)
* [MusDB-HQ](https://zenodo.org/records/3338373) (Human)
* [Music4All](https://sites.google.com/view/contact4music4all) (Human)
* [Million Song Dataset](http://millionsongdataset.com/) (Human)
#### Speech:
* [MSP-Podcast](https://ecs.utdallas.edu/research/researchlabs/msp-lab/MSP-Podcast.html) (Human)
* [JL-Corpus](https://github.com/tli725/JL-Corpus) (Human)
* [MELD](https://github.com/declare-lab/MELD) (Human)
* [Tess](https://www.kaggle.com/datasets/ejlok1/toronto-emotional-speech-set-tess) (Human)
* [OMGEmotion](https://github.com/knowledgetechnologyuhh/OMGEmotionChallenge) (Human)
* [Emov-DB](https://github.com/numediart/EmoV-DB) (Human)
* [LibriSpeech](https://www.openslr.org/12) (Human)
* [SPGISpeech](https://datasets.kensho.com/datasets/spgispeech) (Human)
* [TEDLIUM](https://www.openslr.org/51/) (Human)
* [GigaSpeech](https://github.com/SpeechColab/GigaSpeech) (Human)
* [Common Voice 15](https://huggingface.co/datasets/mozilla-foundation/common_voice_12_0) (Human)
* [VoxPopuli](https://github.com/facebookresearch/voxpopuli) (Human)
* [VoxCeleb2](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox2.html) (Human)
* [Switchboard](https://catalog.ldc.upenn.edu/LDC97S62) (Human)
* [AMI](https://groups.inf.ed.ac.uk/ami/corpus/) (Human)
#### Voice:
* [VoiceAssistant-400K](https://huggingface.co/datasets/gpt-omni/VoiceAssistant-400K) (Automated)
#### Mixed:
* [AudioSkills-XL (ours)](https://huggingface.co/datasets/nvidia/AudioSkills) (Automated)
* [LongAudio-XL (ours)](https://huggingface.co/datasets/nvidia/LongAudio) (Automated)
* [AF-Think (ours)](https://huggingface.co/datasets/nvidia/AF-Think) (Automated)
* [AF-Chat (ours)](https://huggingface.co/datasets/nvidia/AF-Chat) (Automated)
---
### Testing Dataset:
Audio Flamingo 3 is evaluated on the test split of the following datasets.
Data Collection Method: Human (for all datasets noted below)
Labeling Method: See below
* [ClothoAQA](https://zenodo.org/records/6473207) (Human)
* [MusicAVQA](https://gewu-lab.github.io/MUSIC-AVQA/) (Human)
* [Clotho-v2](https://github.com/audio-captioning/clotho-dataset/tree/master) (Human)
* [CochlScene](https://github.com/cochlearai/cochlscene) (Human)
* [NonSpeech7K](https://zenodo.org/records/6967442) (Human)
* [NSynth](https://magenta.tensorflow.org/datasets/nsynth) (Human)
* [AudioCaps](https://github.com/cdjkim/audiocaps) (Human)
* [US8K](https://urbansounddataset.weebly.com/urbansound8k.html) (Human)
* [GTZAN](https://www.tensorflow.org/datasets/catalog/gtzan) (Human)
* [MMAU](https://github.com/Sakshi113/mmau/tree/main) (Human)
* [MMAR](https://arxiv.org/abs/2505.13032) (Human)
* [Audio Entailment](https://github.com/microsoft/AudioEntailment)(Automated)
* [CompA-R-test](https://github.com/Sreyan88/GAMA) (Automated)
* [MuchoMusic](https://huggingface.co/datasets/yongyizang/RUListening) (Automated)
* [Open-AQA](https://github.com/YuanGongND/ltu?tab=readme-ov-file)(Automated)
* [MusicInstruct](https://huggingface.co/datasets/m-a-p/Music-Instruct) (Automated)
* [MusicQA](https://huggingface.co/datasets/mu-llama/MusicQA) (Automated)
* [CMM Hallucination](https://huggingface.co/datasets/DAMO-NLP-SG/CMM) (Human)
* [IEMOCAP](https://sail.usc.edu/iemocap/) (Human)
* [VoiceBench](https://github.com/MatthewCYM/VoiceBench) (Human)
* [OpenAudioBench](https://huggingface.co/datasets/baichuan-inc/OpenAudioBench) (Human)
* [SEED](https://github.com/BytedanceSpeech/seed-tts-eval) (Human)
* [LibriSpeech](https://www.openslr.org/12) (Human)
* [SPGISpeech](https://datasets.kensho.com/datasets/spgispeech) (Human)
* [TEDLIUM](https://www.openslr.org/51/) (Human)
* [GigaSpeech](https://github.com/SpeechColab/GigaSpeech) (Human)
* [Common Voice 15](https://huggingface.co/datasets/mozilla-foundation/common_voice_12_0) (Human)
* [VoxPopuli](https://github.com/facebookresearch/voxpopuli) (Human)
* [LongAudioBench (ours)](https://huggingface.co/datasets/nvidia/LongAudio) (Automated)
* [AF-Chat-test (ours)](https://huggingface.co/datasets/nvidia/AF-Chat) (Human)
---
## Inference:
**Engine:** HuggingFace Transformers
**Test Hardware:** NVIDIA A100 80 GB
---
## Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
---
## Acknowledgements
Built with Qwen, NVILA and the open audio-ML community. |