Qwen3-ASR-API
Pure OpenAI-compatible speech-to-text (ASR) API powered by Qwen3-ASR via vLLM, with a ready-to-run entrypoint for Docker and Unraid.
Platform availability
Available on 1 of 10 platforms
- Unraid
- TrueNAS (not listed)
- Umbrel (not listed)
- ZimaOS (not listed)
- Proxmox (not listed)
- Helm (not listed)
- Mac (not listed)
- Windows (not listed)
- Linux (not listed)
- Any server (not listed)
Health score?A 0-100 score computed nightly from four factors: maintained (40%: recent commits and steady activity), popular (25%: GitHub stars, log scale), easy to install (20%: platforms, install notes, setup guides, Docker image) and light to run (15%: minimum RAM, ARM support). When the source code has been checked, a fifth factor, security check (20%: automated scan for committed secrets, unsafe desktop-app settings, known-vulnerable dependencies and project hygiene; any serious finding caps it at 20), is added. Factors without data are left out and the rest are rescaled.
35/100 Low
- Maintained
- 59last commit 165 days ago
- Popular
- 102 GitHub stars
- Easy to install
- 351 install platform, install notes, Docker image
- Light to run
- 108 GB minimum RAM
Availability: platform listings confirmed today, links checked 18 days ago
- GitHub stars
- 2
- Open issues
- 0
- Last commit
- 2026-06-08
- License
- Apache-2.0
- Activity
- stale
Checked 6 days ago - source: GitHub (hsiang-han/Qwen3-ASR-API)
Resources & Compatibility
Minimum requirements: 8 GB RAM.
Checked 54 days ago - source: Docker Hub image tags and published docs
First-install notes
- Set your API URL and key first Before calling the API, set your NVIDIA NIM invoke URL and put your API key in the Authorization header. If these are missing or wrong, your request will be rejected and you will not get any results. source
- Poll long-running requests If the API returns a 202 status, save the requestId and keep checking the status endpoint until it returns 200. Without polling, you only see that the request is still working and may wrongly think it failed. source
- Name LangChain images correctly If you use LangChain with image inputs, save the image files as image_1.png, image_2.png, and so on in the same folder. If the names are different, LangChain may not send the images you expect. source
- Hardware requirements Requires an NVIDIA GPU with driver 550+ and CUDA 12.8 (Blackwell/Ada). The 0.6B model uses about 2-3 GB of VRAM; the 1.7B model uses about 4-6 GB of VRAM. source
- First run & model switching On first start the model is downloaded from Hugging Face. China-based users can set HF_ENDPOINT to https://hf-mirror.com for faster downloads. Switch between the 0.6B and 1.7B models with the MODEL_ID environment variable. source
Fetched 48 days ago - each note links to its source
Variants
Alternatives
No alternatives collected yet.
Common questions
How do I run this on Unraid without a custom entrypoint?
The image includes an entrypoint that starts qwen-asr-serve automatically, so it can be installed from the Unraid Community Applications template. Configure the model and GPU settings in the template, then start the container.
How do I switch between the 0.6B and 1.7B models?
Set the MODEL_ID environment variable to Qwen/Qwen3-ASR-0.6B or Qwen/Qwen3-ASR-1.7B. Restart the container after changing it.
Where is the model downloaded, and how do I avoid re-downloading on restart?
Mount a persistent directory to /root/.cache/huggingface in the container. The first start downloads the model, and later starts use the cached copy.
What port should I expose, and how do clients call it?
The container listens on port 80 by default, so a common mapping is 8000:80. Clients use the OpenAI-compatible audio transcription endpoint with the model name qwen3-asr.
How do I handle transcriptions longer than the default window?
MAX_MODEL_LEN controls the maximum sequence length for the KV cache and defaults to 8192, which supports about 10 minutes of audio. Raise it for longer audio, or lower it to save VRAM.
Answers sourced from github.com
Community
- Reddit (reddit)
Collected 49 days ago - public community threads