Step-Audio-R1.1

The First Audio Language Model with Test-Time Compute Scaling

English | 简体中文 | 繁體中文 | 日本語

---

🎯 Overview

Step-Audio-R1.1 is a state-of-the-art audio language model that combines real-time responsiveness with strong reasoning capability. It's the first audio model to successfully unlock test-time compute scaling, surpassing Gemini 2.5 Pro on comprehensive audio benchmarks.

✨ Key Features

Feature	Description
🎙️ 5 Processing Modes	ASR, Speech-to-Text, Translation, Summarization, Understanding
⚡ High Concurrency	Process 4× 85-minute audio files simultaneously
🧠 Deep Reasoning	Chain-of-Thought reasoning grounded in acoustic features
🐳 All-in-One Docker	Single container with vLLM + Web UI + API
🌐 Web UI	Beautiful, responsive interface for audio processing
📡 REST API	OpenAI-compatible API for integration
🔧 Long Audio Support	Automatic segmentation for audio >85 minutes

---

🚀 Quick Start

Step 1: Download Model (Required, ~65GB)

# Method 1: Git LFS (Recommended)
git lfs install
git clone https://huggingface.co/stepfun-ai/Step-Audio-R1.1

# Method 2: Hugging Face CLI
pip install huggingface_hub
huggingface-cli download stepfun-ai/Step-Audio-R1.1 --local-dir ./Step-Audio-R1.1

Step 2: Run with Docker

# Pull the all-in-one image
docker pull neosun/step-audio-r1.1:latest

# Run (mount the model directory)
docker run --gpus all \
  -v $(pwd)/Step-Audio-R1.1:/model:ro \
  -p 9100:9100 \
  -p 9101:9999 \
  neosun/step-audio-r1.1:latest

Alternative: Docker Compose

# Clone the repository
git clone https://github.com/neosun100/Step-Audio-R1.1.git
cd Step-Audio-R1.1

# Download model (if not already done)
git lfs install
git clone https://huggingface.co/stepfun-ai/Step-Audio-R1.1

# Start services
docker compose up -d

Access the Web UI at: http://localhost:9100

One-Click Start Script

For convenience, use the auto-start script that finds the model automatically:

# Run from anywhere
./start-step-audio.sh

# Or install globally
sudo cp start-step-audio.sh /usr/local/bin/step-audio
step-audio

The script will:

• Auto-detect model location
• Restart stopped containers
• Pull image if needed
• Wait for service ready

---

📦 Installation

Prerequisites

• GPU: 4× NVIDIA GPUs with ≥40GB VRAM each (tested on L40S/H100/H800)
• Docker: 20.10+ with NVIDIA Container Toolkit
• Storage: ~65GB for model files

Download Model

# Method 1: Git LFS
git lfs install
git clone https://huggingface.co/stepfun-ai/Step-Audio-R1.1

# Method 2: Hugging Face CLI
pip install huggingface_hub
huggingface-cli download stepfun-ai/Step-Audio-R1.1 --local-dir ./Step-Audio-R1.1

Configuration

Create .env file from template:

cp .env.example .env

Variable	Default	Description
WEB_PORT	9100	Web UI port
VLLM_PORT	9101	vLLM API port
MODEL_PATH	./Step-Audio-R1.1	Path to model files
TENSOR_PARALLEL_SIZE	4	Number of GPUs
MAX_NUM_SEQS	4	Max concurrent requests
GPU_MEMORY_UTILIZATION	0.85	GPU memory usage

---

🎮 Usage

Web UI

Navigate to http://localhost:9100 and:

1. Upload an audio file (WAV, MP3, FLAC, M4A)
2. Select processing mode
3. Click "Process" and wait for results

API

# Process audio file
curl -X POST http://localhost:9100/api/process \
  -F "audio=@your_audio.wav" \
  -F "mode=summarize"

# Available modes: asr, s2t, translate, summarize, understand

Python SDK

import requests

with open("audio.wav", "rb") as f:
    response = requests.post(
        "http://localhost:9100/api/process",
        files={"audio": f},
        data={"mode": "understand", "question": "What is discussed?"}
    )
print(response.json()["answer"])

Long Audio Processing

For audio longer than 85 minutes, use the smart processor:

python smart_audio_processor.py input.wav -m summarize -o output.json -p 4

---

📊 Benchmark Results

Processing Modes Comparison (5-minute audio)

Mode	Time	Output	Best For
ASR	52.8s	4,496 chars	Precise transcription
S2T	46.8s	3,713 chars	Structured notes
Translate	51.7s	1,690 chars	Cross-language
Summarize	26.0s	1,637 chars	Quick overview
Understand	29.5s	2,025 chars	Deep analysis

Concurrency Test (4× 85-minute audio)

Mode	Success Rate	Total Time
All modes	20/20 ✅	~250s each

Processing Efficiency

Audio Length	Efficiency (sec/min)
5 min	5-13
30 min	1-3
60-85 min	0.5-0.8

---

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│              Step-Audio-R1.1 Container              │
├─────────────────────────────────────────────────────┤
│  ┌─────────────┐    ┌─────────────────────────────┐ │
│  │   Web UI    │    │         vLLM Server         │ │
│  │  (Flask)    │───▶│  (Qwen2.5 32B + Audio Enc)  │ │
│  │  Port 9100  │    │        Port 9999            │ │
│  └─────────────┘    └─────────────────────────────┘ │
│         │                       │                   │
│         ▼                       ▼                   │
│  ┌─────────────┐    ┌─────────────────────────────┐ │
│  │  REST API   │    │    4× GPU Tensor Parallel   │ │
│  └─────────────┘    └─────────────────────────────┘ │
└─────────────────────────────────────────────────────┘

---

📁 Project Structure

Step-Audio-R1.1/
├── app.py                    # Web UI + API server
├── smart_audio_processor.py  # Long audio processor
├── docker-compose.yml        # All-in-one deployment
├── Dockerfile.allinone       # All-in-one image
├── static/                   # Frontend assets
├── templates/                # HTML templates
├── test_audio/               # Test reports
│   ├── FULL_BENCHMARK_REPORT.md
│   └── CONCURRENCY_TEST_REPORT.md
└── assets/                   # Documentation images

---

🔧 Advanced Configuration

Custom Docker Compose

services:
  step-audio:
    image: neosun/step-audio-r1.1:latest
    runtime: nvidia
    environment:
      - TENSOR_PARALLEL_SIZE=4
      - MAX_NUM_SEQS=4
      - GPU_MEMORY_UTILIZATION=0.85
    volumes:
      - ./Step-Audio-R1.1:/model:ro
    ports:
      - "9100:9100"
      - "9101:9999"

Environment Variables

Variable	Description
TENSOR_PARALLEL_SIZE	Number of GPUs for tensor parallelism
MAX_MODEL_LEN	Maximum context length (default: 65536)
MAX_NUM_SEQS	Maximum concurrent requests
GPU_MEMORY_UTILIZATION	GPU memory fraction to use

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

---

📝 Changelog

v1.1.0 (2026-01-18)

• ✨ All-in-One Docker image with vLLM + Web UI
• ✨ Support for 4× concurrent 85-minute audio processing
• ✨ Smart audio processor for unlimited length audio
• ✨ Comprehensive benchmark reports
• 🐛 Fixed ASR truncation issues for long audio
• 📚 Multi-language documentation

v1.0.0 (2026-01-14)

• 🎉 Initial release of Step-Audio-R1.1
• ✨ 5 processing modes (ASR, S2T, Translate, Summarize, Understand)
• ✨ Web UI and REST API

---

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

---

🙏 Acknowledgments

• StepFun AI for the original Step-Audio-R1 model
• vLLM for the high-performance inference engine

---

⭐ Star History

---

📱 Follow Us

---

Made with ❤️ by NeoSu