Qwen /

Qwen3-ASR-1.7B

An open speech-to-text model that transcribes 52 languages and dialects, with optional timestamps.

1.2k likes
1.3M downloads a month
2.3B parameters
Some experience
Speech

Open on Hugging Face (opens in a new tab)

Qwen3-ASR-1.7B is an automatic speech recognition model from the Qwen team. It is part of a family with a smaller 0.6B sibling. Both are built on the Qwen3-Omni foundation model, handle language identification and transcription, and run in offline or streaming mode. The card says the 1.7B version is state-of-the-art among open-source ASR models and competitive with the strongest proprietary commercial APIs. Those are the authors' own claims from their experiments, so test it on your own audio.

What sets it apart is the toolkit around it. A separate model, Qwen3-ForcedAligner-0.6B, adds timestamps for up to 5 minutes of speech in 11 languages, and the qwen-asr package supports vLLM batch inference, serving and streaming. Honest limits: the card's examples all use a CUDA GPU, streaming works only with the vLLM backend, and streaming does not support batching or timestamps. The card's quickstart shows no CPU-only route, and the 2000x throughput figure belongs to the 0.6B model, not this one.

The card also lists all the supported languages and dialects, so check that yours is included before you build on it.

What you need to run it

2.3B parameters. The weights take about 5 GB at full 16-bit precision (fits a 8 GB Mac or a GPU with 8 GB), or about 2 GB as a 4-bit download, which runs on a 8 GB Mac or a GPU with 8 GB. Long inputs and images need more memory on top.

What it does well

Many languages in one model
Covers 30 languages and 22 Chinese dialects, plus English accents from multiple countries and regions, with language identification built in.
Handles singing and music
Audio types listed include speech, singing voice, and songs with background music, not just clean speech.
Optional timestamps
Pair it with Qwen3-ForcedAligner-0.6B to get word or character level timestamps, for up to 5 minutes of speech in 11 languages.
Streaming and offline with one model
The same model does both modes and can transcribe long audio. Streaming needs the vLLM backend.
Full tooling included
A pip package, a Docker image, a Gradio web demo, a streaming demo and an OpenAI-style vLLM server are all described in the card.

What you could build

Transcribe lecture recordings or interviews and get the detected language along with the text.
Build subtitles for a video by pairing the model with the forced aligner to get timestamps.
Run a local speech-to-text server and call it from a small app using the OpenAI-style transcription API.
Try the Gradio web demo to record from a microphone and watch live streaming transcription.
Compare the 1.7B and 0.6B models on your own recordings to see the accuracy and speed trade-off.

Get started

Create a clean Python 3.12 environment, install the qwen-asr package, then load Qwen/Qwen3-ASR-1.7B with the Python example in the card and call transcribe on an audio file or URL. The weights download automatically on first load.

conda create -n qwen3-asr python=3.12 -y

conda activate qwen3-asr

pip install -U qwen-asr

pip install -U qwen-asr[vllm]

Commands are copied from the model card. Read it on Hugging Face for the full setup.