orukqwen-asr-1.7b

orukqwen-asr-1.7b

Turns English speech into text. It is Qwen3-ASR-1.7B post-trained by reinforcement learning on word error rate, for developers and speech researchers who need English transcription.

Made by oruk on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: orukqwen-asr-1.7b is an English speech recognition model from oruk, with 2.04B parameters. It starts from Qwen3-ASR-1.7B and changes only the weights, using reinforcement learning that scores each transcript by its word error rate. The card calls it a research artifact.

Who it suits: It suits developers who need English transcription and researchers who study how speech models are trained.

Industry | Who uses it | How they use it | Example
Speech research | Researchers working on speech recognition | Study the reinforcement learning and weight blending method | A researcher compares its word error rate with the base model on their own audio
Software development | Developers adding English transcription | Swap these weights into a pipeline built for the base model | A developer replaces the base weights and keeps the same inference code
Transcription work | Teams that turn English recordings into text | Run recordings through the model and review the output | A team transcribes interview audio

Anyone who needs a language other than English can skip it, because the record lists English only.

Where to use it: It is a set of model files that you run with Python and the transformers library.

• Windows, macOS or Linux: download the files with Python and the Hugging Face tools
• The transformers library: load it through the same layout as Qwen/Qwen3-ASR-1.7B-hf
• Speech research setups: compare it with the base model

Use case | Who | What they get back
Transcribe English speech | Developers | A text transcript of the audio
Benchmark a speech model | Researchers | A word error rate to compare with the base model
Study an RL recipe | Speech researchers | A worked example of word-level rewards and weight blending

How to set it up: You need a few things first.

• Python and the huggingface_hub package for the download
• The transformers library, which the record lists for this model
• About 4.1 GB of disk for the BF16 weights

1. Install the Hugging Face hub package.
2. Download the model files.
3. Load them the way you would load Qwen/Qwen3-ASR-1.7B-hf, since only the weights differ.

pip install -U huggingface_hub
hf download oruk/orukqwen-asr-1.7b

How to use it: Feed English audio to the model the way you would to its base. The card says the architecture, tokenizer, prompt format and inference code match Qwen/Qwen3-ASR-1.7B-hf.

Tool | What it does
hf download | Fetches the model files to your computer
huggingface_hub | The Python package that provides the hf command
Qwen/Qwen3-ASR-1.7B-hf | The base layout whose tokenizer, prompt format and inference code this model reuses

Example: You have an English recording of a team meeting and ask the model to transcribe it. A text transcript comes back. The model was trained to score well on word error rate, so the transcript is judged against the words that were spoken.

For newcomers and experts: This model turns English speech into text. It is an existing speech model, retrained so that it makes fewer word-level mistakes. Your first small step is to download the files and run a short recording through it as you would with the base model.

The training samples 64 transcripts per recording and rewards each by one minus its word error rate, capped at one, with no learned judge. A word-level alignment zeroes the advantage on wrongly transcribed words inside transcripts that scored well, so those errors are not reinforced. Three related runs are averaged evenly, then mixed half and half with the base weights in the WiSE-FT manner. The card is open that this objective is biased and differs from the standard one, trading that for lower variance.

Start from Qwen3-ASR-1.7B
Train with RL and word-level credit assignment
Average three related runs into a soup
Blend the soup half and half with the base

Good to know:

• Training used one NVIDIA A100 40GB card at a time, with a pool of 54.6 hours of English audio that has transcripts.
• A public H200 reproduction on the Open ASR Leaderboard gave a nine-set mean WER of 4.76 and an aggregate RTFx of 718.84. The maintainer's private-set evaluation is still pending.
• Training updated 1.44B parameters out of 2.04B. That covers the language decoder, the last two audio layers and the audio projector. The embeddings, the lm_head and the remaining encoder layers were left untouched, so they match the base.

Title: Compute used for training
RL training in total: 14 GPU-hours
Data selection pass at most: 7.5 GPU-hours

• License: Apache 2.0, the same as its base model.
• Cost to run: none named, no API key, account or paid service is mentioned.
• Data it reaches: it downloads files from Hugging Face and reads the audio you give it.
• Not for: languages other than English, since the record lists only English.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 2B
• Weights: about 4.1 GB in BF16; about 1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 222
• Likes: 6
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3-ASR-1.7B
• Tags: transformers, safetensors, qwen3_asr, automatic-speech-recognition, speech, audio, qwen3-asr, reinforcement-learning, gspo, model-soup, wise-ft, research

Links

• Model: https://huggingface.co/oruk/orukqwen-asr-1.7b
• Files: https://huggingface.co/oruk/orukqwen-asr-1.7b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download oruk/orukqwen-asr-1.7b

Then follow the usage in its model card. Nothing in this kit runs by itself.