MOSS-Audio

MOSS-Audio

English | 简体中文

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.

Start here: Run locally · Gradio app · Model weights · Fine-tuning · Evaluation

News

• 2026.6.1: We have released the MOSS-Audio Technical Report on arXiv.
• 2026.4.20: We have added the MOSS-Audio fine-tuning code and documentation. See finetune/FINETUNE.md for LoRA and full-parameter training examples.
• 2026.4.13: 🎉🎉🎉 We have released MOSS-Audio. Blog coming soon!

Contents

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

Introduction

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

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

Model	Audio Encoder	LLM Backbone	Total Size	Hugging Face	ModelScope
MOSS-Audio-4B-Instruct	MOSS-Audio-Encoder	Qwen3-4B	~4.6B
MOSS-Audio-4B-Thinking	MOSS-Audio-Encoder	Qwen3-4B	~4.6B
MOSS-Audio-8B-Instruct	MOSS-Audio-Encoder	Qwen3-8B	~8.6B
MOSS-Audio-8B-Thinking	MOSS-Audio-Encoder	Qwen3-8B	~8.6B

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 71.08, with 77.33 on MMAU, 64.92 on MMAU-Pro, 66.53 on MMAR, and 75.52 on MMSU, outperforming all 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↑)

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

ASR

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:

hf  download OpenMOSS-Team/MOSS-Audio-4B-Instruct --local-dir ./weights/MOSS-Audio-4B-Instruct 
hf  download OpenMOSS-Team/MOSS-Audio-4B-Thinking --local-dir ./weights/MOSS-Audio-4B-Thinking 
hf  download OpenMOSS-Team/MOSS-Audio-8B-Instruct --local-dir ./weights/MOSS-Audio-8B-Instruct 
hf  download OpenMOSS-Team/MOSS-Audio-8B-Thinking --local-dir ./weights/MOSS-Audio-8B-Thinking 

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.

Fine-tuning

We now provide an official fine-tuning script in finetune/finetune.py, with full instructions in finetune/FINETUNE.md.

Install the extra dependencies needed for training:

pip install librosa peft

Minimal example for LoRA fine-tuning:

accelerate launch finetune/finetune.py \
    --model_dir ./weights/MOSS-Audio-4B-Instruct \
    --data_path train.jsonl \
    --output_dir ./output/lora \
    --use_lora \
    --bf16

The training data should be a JSONL file containing audio-text conversations. For data format, supported arguments, multi-GPU examples, and full-parameter fine-tuning, see finetune/FINETUNE.md.

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.

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{yang2026mossaudiotechnicalreport,
      title={MOSS-Audio Technical Report}, 
      author={Chen Yang and Chufan Yu and Hanfu Chen and Jie Zhu and Jingqi Chen and Ke Chen and Wenxuan Wang and Yang Wang and Yaozhou Jiang and Yi Jiang and Zhengyuan Lin and Ziqi Chen and Zhaoye Fei and Chenghao Liu and Jun Zhan and Kang Yu and Kexin Huang and Mingshu Chen and Qinyuan Cheng and Ruixiao Li and Shimin Li and Songlin Wang and Yang Gao and Yiyang Zhang and Xipeng Qiu},
      year={2026},
      eprint={2606.01802},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2606.01802}, 
}

Star History