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.