Skip to content
Audio

Qwen3-ASR

streaming speech recognition for RK3588

2 min readupdated

This model delivers step-by-step Automatic Speech Recognition (ASR) with real-time text output and language detection. The build is optimized for RK182x and RK3588 hardware platforms using the RKLLM and RKNN engines. Built on the 0.6B Qwen3-ASR architecture, it processes audio streams in one-second chunks. It is a great fit for embedded systems and smart device developers working with Rockchip SoCs who need local, streaming voice processing without sending data to the cloud.

Key features

  • Support for 52 languages and dialects (including numerous regional varieties and dialects).
  • Streaming inference: text updates every 1000 ms with token feedback.
  • Low memory footprint: requires around 1600 MB of VRAM or RAM for weights and context.
  • Ready-to-use manifest for integration via the wdnwebui web interface.

What you can do

  • Hook up a microphone to an RK3588 single-board computer for real-time voice input and transcription.
  • Transcribe audio tracks from uploaded videos or audio recordings directly on the device.
  • Embed a local speech recognition service into offline smart speakers and robotics systems.

Example

Download the repository with the weights and manifest for local deployment:

hf download wardner/Qwen3-ASR --local-dir ./Qwen3-ASR
Read moreDetails
Compute unit: RK182x (target board RK3588)
Runtime: RKLLM + RKNN
Total weights size: 1.05 GB (encoder, LLM, embeddings, tokenizer)
Limitations: maximum audio duration for streaming mode is capped; input audio is automatically resampled to 16 kHz mono. License: Apache 2.0 (for the base model).
Discuss Telegram 0 334

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.3 GB · ready files0.8 GB
  • FP8 / INT8recommendedweights 0.1 GB0.7 GB
  • 4-bitweights 0.1 GB0.6 GB
  • GGUFweights 0.0 GB · ready files0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # wardner/Qwen3-ASR
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download wardner/Qwen3-ASR --local-dir ~/models/Qwen3-ASR
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3-ASR
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-ASR (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

wardner/Qwen3-ASR
Author
wardner
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

3340

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments