qwen3_asr_1.7b_ko

Use it to transcribe Korean speech from audio files, with language detection or Korean specified explicitly.

Availability

Available from the TeamUNIVA Hugging Face repository

Pricing model

Not established

Deployment

Load with the qwen-asr package; example uses CUDA

Privacy

Not established

Licence

Apache-2.0

• The model is fine-tuned for Korean automatic speech recognition. ^[1]

  This model is a fine-tune of `Qwen/Qwen3-ASR-1.7B` on Korean speech data for automatic speech recognition (ASR), and the successor to [`TeamUNIVA/qwen3_asr_1.7b_ko_beta`](https://huggingface.co/TeamUNIVA/qwen3_asr_1.7b_ko_beta).

• It achieves a Korean total CER of 0.0483 on the reported benchmarks. ^[1]
• Its English LibriSpeech WER is 0.0277, compared with 0.0250 for the base model. ^[1]

  It achieves lower error rates than both the base model and the beta model across Korean benchmarks, with a small degradation on English (LibriSpeech, +0.27 %p WER vs.

• The model can be loaded with Qwen3ASRModel.from_pretrained using the TeamUNIVA/qwen3_asr_1.7b_ko identifier. ^[1]

  Starting from `TeamUNIVA/qwen3_asr_1.7b_ko_beta` (itself an SFT-only fine-tune of `Qwen/Qwen3-ASR-1.7B`), the model was trained with full-parameter updates in four sequential stages, using [ms-swift](https://github.com/modelscope/ms-swift):

• The model’s performance on other languages may be similar to the base model or require separate validation. ^[1]

  - This model was fine-tuned specifically for Korean speech recognition; performance on other languages may be similar to the base model or require separate validation.

• An open model for tool-using agents, plus a checkpoint to fine-tune Daily brief 2026-10-13