MOSS-Audio-8B-Thinking

MOSS-Audio

<p align="center"> <img src="./assets/moss-audio-logo.png" width="55%" /> </p> <div align="center"> <a href="https://huggingface.co/collections/OpenMOSS-Team/moss-audio"><img src="https://img.shields.io/badge/Huggingface-Models-orange?logo=huggingface&amp"></a> <img src="https://img.shields.io/badge/Blog-Coming_Soon-blue?logo=internet-explorer&amp"> <img src="https://img.shields.io/badge/Arxiv-Coming_Soon-red?logo=Arxiv&amp">

<a href="https://x.com/Open_MOSS"><img src="https://img.shields.io/badge/Twitter-Follow-black?logo=x&amp"></a> <a href="https://discord.gg/Xf3aXddCjc"><img src="https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&amp"></a> <a href="./assets/wechat.jpg"><img src="https://img.shields.io/badge/WeChat-Join-07C160?logo=wechat&amp;logoColor=white" alt="WeChat"></a>

</div>

MOSS-Audio is an open-source audio understanding model from MOSI.AI, the OpenMOSS team, and Shanghai Innovation Institute. It performs unified modeling over complex real-world audio, supporting speech understanding, environmental sound understanding, music understanding, audio captioning, time-aware QA, and complex reasoning. In this release, we provide four models: MOSS-Audio-4B-Instruct, MOSS-Audio-4B-Thinking, MOSS-Audio-8B-Instruct, and MOSS-Audio-8B-Thinking. The Instruct variants are optimized for direct instruction following, while the Thinking variants provide stronger chain-of-thought reasoning capabilities.

News

• 2026.4.13: 🎉🎉🎉 We have released MOSS-Audio. Blog and paper coming soon!

Contents

• Introduction
• Model Architecture
  • DeepStack Cross-Layer Feature Injection
  • Time-Aware Representation
• Released Models
• Evaluation
• Quickstart
  • Environment Setup
  • Basic Usage
  • Gradio App
  • SGLang Serving
• More Information
• Citation

Introduction

<p align="center"> <img src="./assets/moss-audio-image.png" width="95%" /> </p>

Understanding audio requires more than simply transcribing words — it demands the ability to perceive acoustic cues, recognize speakers and emotions, interpret environmental sounds, reason over temporal context, and handle complex multi-step inference. MOSS-Audio is built to unify these capabilities within a single model.

• Speech & Content Understanding: Accurately recognizes and transcribes spoken content from audio inputs, producing clean and well-structured text outputs. Supports both word-level and sentence-level timestamp alignment.
• Speaker, Emotion & Event Analysis: Identifies speaker characteristics, analyzes emotional states based on tone, timbre, and context, and detects key acoustic events within the audio.
• Scene & Sound Cue Extraction: Extracts meaningful cues from background sounds, environmental noise, music, and non-speech signals to infer scene context and atmosphere.
• Music Understanding: Analyzes musical style, emotional progression, instrumentation, and salient acoustic features in music segments.
• Audio Question Answering & Summarization: Answers questions and generates summaries about speech, podcasts, meetings, interviews, and environmental recordings, helping users efficiently extract key information.
• Time-Aware QA: Supports time-aware questions, including word-level and sentence-level timestamp ASR.
• Complex Reasoning: Performs multi-hop reasoning over audio content, powered by chain-of-thought training and reinforcement learning.

Model Architecture

<p align="center"> <img src="./assets/arc.png" width="95%" /> </p>

MOSS-Audio follows a modular design comprising three components: an audio encoder, a modality adapter, and a large language model. Raw audio is first encoded by MOSS-Audio-Encoder into continuous temporal representations at 12.5 Hz, which are then projected into the language model's embedding space through the adapter and finally consumed by the LLM for auto-regressive text generation.

Rather than relying on off-the-shelf audio frontends, we train a dedicated encoder from scratch to obtain more robust speech representations, tighter temporal alignment, and better extensibility across acoustic domains.

DeepStack Cross-Layer Feature Injection

Using only the encoder's top-layer features tends to lose low-level prosody, transient events, and local time-frequency structure. To address this, we design a DeepStack-inspired cross-layer injection module between the encoder and the language model: in addition to the encoder's final-layer output, features from earlier and intermediate layers are selected, independently projected, and injected into the language model's early layers, preserving multi-granularity information from low-level acoustic details to high-level semantic abstractions.

This design is especially well-suited for audio understanding tasks, as it helps retain rhythm, timbre, transients, and background structure — information that a single high-level representation cannot fully capture.

Time-Aware Representation

Time is a critical dimension in audio understanding. To enhance explicit temporal awareness, we adopt a time-marker insertion strategy during pretraining: explicit time tokens are inserted between audio frame representations at fixed time intervals to indicate temporal positions. This design enables the model to learn "what happened when" within a unified text generation framework, naturally supporting timestamp ASR, event localization, time-based QA, and long-audio retrospection.

Released Models

More model families, sizes, and variants will be released in the future. Stay tuned!

Evaluation

We evaluate MOSS-Audio on a comprehensive set of audio understanding benchmarks. Key results:

• General Audio Understanding: MOSS-Audio-8B-Thinking achieves an average accuracy of 70.80, outperforming all of the open-source models.
• Speech Captioning: MOSS-Audio-Instruct variants lead across 11 out of 13 fine-grained speech description dimensions, with MOSS-Audio-8B-Instruct achieving the best overall average score (3.7252).
• ASR: On a diverse ASR benchmark suite spanning 12 evaluation dimensions, MOSS-Audio achieves the lowest overall CER (11.30), with particular strength in health-condition, code-switching, dialect, singing, and non-speech scenarios.
• Timestamp ASR: MOSS-Audio-8B-Instruct achieves 35.77 AAS on AISHELL-1 and 131.61 AAS on LibriSpeech, dramatically outperforming Qwen3-Omni (833.66) and Gemini-3.1-Pro (708.24) in timestamp asr accuracy.

General Audio Understanding (Accuracy↑)

<table> <thead> <tr> <th>Model</th> <th>Model Size</th> <th>MMAU</th> <th>MMAU-Pro</th> <th>MMAR</th> <th>MMSU</th> <th>Avg</th> </tr> </thead> <tbody> <tr><td colspan="7"><em><strong>Open Source (small)</strong></em></td></tr> <tr> <td>Kimi-Audio</td><td>7B</td><td>72.41</td><td>56.58</td><td>60.82</td><td>54.74</td><td>61.14</td> </tr> <tr> <td>Qwen2.5-Omni</td><td>7B</td><td>65.60</td><td>52.20</td><td>56.70</td><td>61.32</td><td>58.96</td> </tr> <tr> <td>Audio Flamingo 3</td><td>7B</td><td>61.23</td><td>51.70</td><td>57.96</td><td>60.04</td><td>57.73</td> </tr> <tr> <td>MiMo-Audio-7B</td><td>7B</td><td>74.90</td><td>53.35</td><td>61.70</td><td>61.94</td><td>62.97</td> </tr> <tr> <td>MiniCPM-o-4.5</td><td>9B</td><td>70.97</td><td>39.65</td><td>55.75</td><td>60.96</td><td>56.83</td> </tr> <tr> <td><strong>MOSS-Audio-4B-Instruct</strong></td><td><strong>4B</strong></td><td>75.79</td><td>58.16</td><td>59.68</td><td>59.68</td><td>64.04</td> </tr> <tr> <td><strong>MOSS-Audio-4B-Thinking</strong></td><td><strong>4B</strong></td><td><strong>77.64</strong></td><td>60.75</td><td>63.91</td><td>71.20</td><td>68.37</td> </tr> <tr> <td><strong>MOSS-Audio-8B-Instruct</strong></td><td><strong>8B</strong></td><td>77.03</td><td>57.48</td><td>64.42</td><td>66.36</td><td>66.32</td> </tr> <tr> <td><strong>MOSS-Audio-8B-Thinking</strong></td><td><strong>8B</strong></td><td>77.33</td><td><strong>64.92</strong></td><td><strong>66.53</strong></td><td><strong>75.52</strong></td><td><strong>71.08</strong></td> </tr> <tr><td colspan="7"><em><strong>Open Source (large)</strong></em></td></tr> <tr> <td>Qwen3-Omni-30B-A3B-Instruct</td><td>30B</td><td>75.00</td><td><strong>61.22</strong></td><td>66.40</td><td>69.00</td><td>67.91</td> </tr> <tr> <td>Step-Audio-R1.1</td><td>33B</td><td>72.18</td><td>60.80</td><td>68.75</td><td>64.18</td><td>66.48</td> </tr> <tr> <td>Step-Audio-R1</td><td>33B</td><td><strong>78.67</strong></td><td>59.68</td><td><strong>69.15</strong></td><td><strong>75.18</strong></td><td><strong>70.67</strong></td> </tr> <tr><td colspan="7"><em><strong>Closed Source</strong></em></td></tr> <tr> <td>GPT4o-Audio</td><td>-</td><td>65.66</td><td>52.30</td><td>59.78</td><td>58.76</td><td>59.13</td> </tr> <tr> <td>Gemini-3-Pro</td><td>-</td><td>80.15</td><td>68.28</td><td>81.73</td><td>81.28</td><td>77.86</td> </tr> <tr> <td>Gemini-3.1-Pro</td><td>-</td><td><strong>81.10</strong></td><td><strong>73.47</strong></td><td><strong>83.70</strong></td><td><strong>81.30</strong></td><td><strong>79.89</strong></td> </tr> </tbody> </table>

Speech Captioning (LLM-as-a-Judge Score↑)

<p align="center"> <img src="./assets/speech_caption_radar.png" width="70%" /> </p> <details> <summary><strong>Speech Captioning (click to expand)</strong></summary> </details>

ASR

<details> <summary><strong>Detailed ASR Results (click to expand)</strong></summary> <table> <tr> <th rowspan="2">Model</th> <th colspan="3">Acoustic Environment (Clean)</th> <th colspan="1">Acoustic Environment (Noisy)</th> <th colspan="1">Acoustic Characteristics: Whisper</th> <th colspan="1">Acoustic Characteristics: Far-Field / Near-Field</th> <th colspan="1">Multi-Speaker</th> <th colspan="2">Age</th> <th colspan="2">Health Condition</th> <th colspan="2">Semantic Content</th> <th colspan="3">Code-Switching</th> <th colspan="2">Dialect</th> <th colspan="2">Singing</th> <th colspan="1">Non-Speech Vocalizations</th> </tr> <tr> <th>AISHELL-1<br><em>test</em></th> <th>AISHELL-2<br><em>Android | IOS | Mic</em></th> <th>THCHS-30<br><em>test</em></th> <th>MAGICDATA-READ<br><em>test</em></th> <th>AISHELL6-Whisper<br><em>normal | whisper</em></th> <th>AliMeeting<br><em>Test_Ali_far | Test_Ali_near</em></th> <th>AISHELL-4<br><em>test</em></th> <th>SeniorTalk<br><em>sentence</em></th> <th>ChildMandarin<br><em>test</em></th> <th>AISHELL-6A<br><em>mild | moderate | severe | StutteringSpeech</em></th> <th>AISHELL_6B<br><em>LRDWWS | Uncontrol</em></th> <th>WenetSpeech<br><em>test-meeting</em></th> <th>Fleurs<br><em>cmn_hans_cn</em></th> <th>CS-Dialogue<br><em>test</em></th> <th>TALCS<br><em>test</em></th> <th>ASCEND<br><em>test</em></th> <th>KeSpeech<br><em>test</em></th> <th>WSYue-ASR-eval<br><em>short</em></th> <th>MIR-1K<br><em>test</em></th> <th>openc-pop<br><em>test</em></th> <th>MNV_17</th> </tr> <tr> <td>Paraformer-Large</td> <td>1.98</td> <td>3.28 | 3.21 | 3.00</td> <td>4.07</td> <td>4.67</td> <td>1.11 | 8.92</td> <td><strong>25.64</strong> | 9.27</td> <td>20.33</td> <td>17.31</td> <td>12.60</td> <td>6.98 | 9.30 | 13.34 | 10.74</td> <td>47.59 | 45.08</td> <td>7.88</td> <td>6.40</td> <td>10.64</td> <td>10.77</td> <td>16.55</td> <td>11.48</td> <td>75.42</td> <td>57.70</td> <td>6.98</td> <td>4.95</td> </tr> <tr> <td>GLM-ASR-Nano</td> <td>2.89</td> <td>3.75 | 3.73 | 3.78</td> <td>4.23</td> <td>5.02</td> <td>0.83 | 9.06</td> <td>40.27 | 14.76</td> <td>28.02</td> <td>20.33</td> <td>14.06</td> <td>8.74 | 12.11 | 14.38 | 12.29</td> <td>50.34 | 49.09</td> <td>9.70</td> <td>4.94</td> <td>11.06</td> <td>11.07</td> <td>13.50</td> <td>9.72</td> <td>35.07</td> <td>95.87</td> <td>8.03</td> <td>4.65</td> </tr> <tr> <td>Fun-ASR-Nano</td> <td>2.16</td> <td>3.04 | 2.99 | 3.07</td> <td>3.65</td> <td>3.46</td> <td>0.81 | 6.76</td> <td>27.21 | 9.55</td> <td>19.82</td> <td>16.96</td> <td>12.94</td> <td>6.60 | <strong>8.81</strong> | 12.98 | 10.30</td> <td>47.42 | 45.84</td> <td>7.39</td> <td><strong>4.76</strong></td> <td>10.47</td> <td><strong>8.09</strong></td> <td>15.13</td> <td>7.43</td> <td>8.17</td> <td>35.85</td> <td>2.84</td> <td>4.76</td> </tr> <tr> <td>SenseVoice-Small</td> <td>3.23</td> <td>4.16 | 4.02 | 3.96</td> <td>5.26</td> <td>4.93</td> <td>1.25 | 9.88</td> <td>37.01 | 16.31</td> <td>24.06</td> <td>21.07</td> <td>14.18</td> <td>7.62 | 9.85 | 14.39 | 11.47</td> <td>52.92 | 47.97</td> <td>8.35</td> <td>6.75</td> <td>12.81</td> <td>10.52</td> <td>18.38</td> <td>10.45</td> <td><strong>7.34</strong></td> <td>39.51</td> <td>8.07</td> <td>4.92</td> </tr> <tr> <td>Kimi-Audio-7B-Instruct</td> <td><strong>0.79</strong></td> <td>2.91 | 3.03 | 2.88</td> <td><strong>1.39</strong></td> <td><strong>2.15</strong></td> <td>0.69 | 4.63</td> <td>28.22 | 13.82</td> <td>20.61</td> <td>19.70</td> <td>13.79</td> <td>7.00 | 9.34 | 12.56 | 10.75</td> <td>44.44 | 42.57</td> <td>7.15</td> <td>5.10</td> <td>14.56</td> <td>12.74</td> <td>21.83</td> <td><strong>5.51</strong></td> <td>53.17</td> <td>38.35</td> <td>5.17</td> <td>4.68</td> </tr> <tr> <td>Qwen2.5-Omni-3B</td> <td>1.51</td> <td>3.10 | 2.94 | 2.93</td> <td>3.32</td> <td>3.56</td> <td>0.82 | 7.82</td> <td>32.14 | 12.16</td> <td>22.91</td> <td>17.38</td> <td>12.96</td> <td>6.87 | 10.55 | 14.57 | 11.33</td> <td>54.54 | 50.03</td> <td>9.04</td> <td>5.45</td> <td>10.78</td> <td>10.94</td> <td>13.25</td> <td>7.67</td> <td>60.06</td> <td>45.00</td> <td>3.47</td> <td>5.54</td> </tr> <tr> <td>Qwen2.5-Omni-7B</td> <td>1.16</td> <td>2.88 | 2.77 | 2.73</td> <td>3.06</td> <td>3.16</td> <td>0.71 | 6.57</td> <td>32.03 | 18.73</td> <td>21.01</td> <td>19.96</td> <td>12.29</td> <td>7.27 | 10.94 | 12.92 | 10.53</td> <td>51.99 | 49.45</td> <td>8.43</td> <td>5.13</td> <td>14.02</td> <td>10.46</td> <td>14.42</td> <td>6.40</td> <td>57.43</td> <td>42.62</td> <td>2.75</td> <td>4.56</td> </tr> <tr> <td>Qwen3-Omni-30B-A3B-Instruct</td> <td>0.95</td> <td><strong>2.70</strong> | <strong>2.72</strong> | <strong>2.57</strong></td> <td>2.21</td> <td>2.47</td> <td><strong>0.59</strong> | <strong>3.22</strong></td> <td>25.72 | <strong>8.44</strong></td> <td><strong>18.15</strong></td> <td><strong>14.13</strong></td> <td><strong>8.79</strong></td> <td>6.20 | 8.88 | 11.59 | 10.25</td> <td>45.80 | 41.65</td> <td><strong>6.64</strong></td> <td>4.84</td> <td>12.94</td> <td>8.33</td> <td><strong>12.64</strong></td> <td>5.87</td> <td>25.39</td> <td>30.81</td> <td><strong>1.21</strong></td> <td>4.73</td> </tr> <tr> <td><strong>MOSS-Audio-4B-Instruct</strong></td> <td>2.26</td> <td>3.22 | 3.20 | 3.33</td> <td>3.53</td> <td>3.72</td> <td>0.73 | 5.86</td> <td>27.27 | 9.68</td> <td>20.33</td> <td>16.93</td> <td>13.25</td> <td>6.36 | 9.77 | 12.68 | 10.28</td> <td>43.35 | 44.25</td> <td>8.17</td> <td>8.13</td> <td>9.14</td> <td>8.37</td> <td>12.83</td> <td>14.65</td> <td>9.04</td> <td>18.47</td> <td>3.10</td> <td><strong>4.01</strong></td> </tr> <tr> <td><strong>MOSS-Audio-8B-Instruct</strong></td> <td>1.82</td> <td>2.97 | 2.95 | 2.91</td> <td>2.82</td> <td>3.20</td> <td>0.69 | 4.80</td> <td>36.82 | 11.25</td> <td>24.36</td> <td>17.42</td> <td>13.10</td> <td><strong>5.84</strong> | 8.94 | <strong>11.52</strong> | <strong>9.72</strong></td> <td><strong>39.76</strong> | <strong>39.27</strong></td> <td>7.86</td> <td>7.52</td> <td><strong>9.07</strong></td> <td>8.22</td> <td>13.26</td> <td>9.18</td> <td>8.33</td> <td><strong>17.24</strong></td> <td>2.39</td> <td>4.31</td> </tr> </table> </details>

Timestamp ASR (AAS↓)

Quickstart

Environment Setup

We recommend Python 3.12 with a clean Conda environment. The commands below are enough for local inference.

Recommended setup

git clone https://github.com/OpenMOSS/MOSS-Audio.git
cd MOSS-Audio

conda create -n moss-audio python=3.12 -y
conda activate moss-audio

conda install -c conda-forge "ffmpeg=7" -y
pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[torch-runtime]"

Optional: FlashAttention 2

If your GPU supports FlashAttention 2, you can replace the last install command with:

pip install --extra-index-url https://download.pytorch.org/whl/cu128 -e ".[torch-runtime,flash-attn]"

Basic Usage

Download the model first:

huggingface-cli download OpenMOSS-Team/MOSS-Audio --local-dir ./weights/MOSS-Audio
huggingface-cli download OpenMOSS-Team/MOSS-Audio-Instruct --local-dir ./weights/MOSS-Audio-Instruct

Then edit MODEL_PATH / AUDIO_PATH in infer.py as needed, and run:

python infer.py

The default prompt in infer.py is Describe this audio. You can directly edit that line if you want to try transcription, audio QA, or speech captioning.

Gradio App

Start the Gradio demo with:

python app.py

SGLang Serving

If you want to serve MOSS-Audio with SGLang, see the full guide in moss_audio_usage_guide.md.

The shortest setup is:

git clone -b moss-audio https://github.com/OpenMOSS/sglang.git
cd sglang
pip install -e "python[all]"
pip install nvidia-cudnn-cu12==9.16.0.29
cd ..
sglang serve --model-path ./weights/MOSS-Audio --trust-remote-code

If you use the default torch==2.9.1+cu128 runtime, installing nvidia-cudnn-cu12==9.16.0.29 is recommended before starting sglang serve.

<a id="more-information"></a>

More Information

• MOSI.AI: https://mosi.cn
• OpenMOSS: https://www.open-moss.com

LICENSE

Models in MOSS-Audio are licensed under the Apache License 2.0.

Citation

@misc{mossaudio2026,
      title={MOSS-Audio Technical Report},
      author={OpenMOSS Team},
      year={2026},
      howpublished={\url{https://github.com/OpenMOSS/MOSS-Audio}},
      note={GitHub repository}
}