# Qwen3-ASR-API

Qwen3-ASR-API is a self-hosted ai app. Pure OpenAI-compatible speech-to-text (ASR) API powered by Qwen3-ASR via vLLM, with a ready-to-run entrypoint for Docker and Unraid. It is open source and self-hosted, installable on Unraid.

Canonical page: https://ourlocalhost.com/app/3248/qwen3-asr-api

## Description

Pure OpenAI-compatible speech-to-text (ASR) API powered by Qwen3-ASR via vLLM, with a ready-to-run entrypoint for Docker and Unraid. It transcribes Chinese (22 dialects), English, and 50+ other languages, and lets you switch between the 0.6B and 1.7B models via environment variables. A lean, model-only service with no NGINX or extra services, aimed at self-hosters who want a single OpenAI-compatible transcription endpoint.

## Platform availability

Available on 1 of 10 platforms: Unraid.

## Health score

35/100 (Low), based on 4 of 4 factors.
- Maintained: 59 (last commit 165 days ago)
- Popular: 10 (2 GitHub stars)
- Easy to install: 35 (1 install platform, install notes, Docker image)
- Light to run: 10 (8 GB minimum RAM)

## Links

- Unraid: https://ca.unraid.net/apps/qwen3-asr-api-0ahfwzz104qmd7?q=Qwen3-ASR-API

## 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.

---
Source: Our Local Host (https://ourlocalhost.com/app/3248/qwen3-asr-api), the self-hosted and open-source app directory.
