Qwen3-ASR-1.7B Albanian

Kushtrim/Qwen3-ASR-1.7B-Albanian is an Albanian automatic speech recognition model fine-tuned from Qwen/Qwen3-ASR-1.7B.

The released weights originate from training step 157,772, at approximately one recorded training epoch.

The training collection contains a couple thousand well-curated hours of Albanian audio and transcripts.

Model details

Evaluation loss is reported only as a training diagnostic; it should not be interpreted as WER or as a production-quality benchmark.

Usage

pip install -U qwen-asr

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "Kushtrim/Qwen3-ASR-1.7B-Albanian",
    dtype=torch.bfloat16,
    device_map="cuda:0",
    max_inference_batch_size=32,
    max_new_tokens=256,
)

results = model.transcribe(
    audio="path/to/audio.wav",
    language="Albanian",
)

print(results[0].text)

For automatic language identification, set language=None.

Intended use

• Albanian speech transcription and ASR research.
• General-purpose Albanian speech-to-text experiments.
• Continued fine-tuning or adaptation to a specific acoustic domain.

Limitations

• No standardized WER/CER result is currently documented for this checkpoint.
• Quality may vary across Albanian dialects, speakers, noise levels, and specialized terminology.
• ASR output may omit or hallucinate words. Human verification is required for consequential applications.
• Evaluate on a representative held-out dataset before deployment.

Provenance

The architecture and training statistics above were recovered from config.json, preprocessor_config.json, and trainer_state.json. Dataset composition, preprocessing, and benchmark methodology should be added when documented. These weights were promoted from the training snapshot originally stored as Kushtrim/Qwen3-ASR-1.7B-Shqip-157772.

Development and attribution

• Original architecture and base weights: Qwen Team, through Qwen/Qwen3-ASR-1.7B
• Albanian adaptation, fine-tuning, training-data curation, checkpoint preparation, packaging, and release: Kushtrim Visoka
• Model card and ongoing maintenance: Kushtrim Visoka

Acknowledgements

This work builds on Qwen3-ASR by the Qwen team. The original architecture and base weights remain the work of their respective authors. This repository documents the Albanian fine-tuning and release work carried out by Kushtrim Visoka and does not claim authorship of the original Qwen3-ASR model.

Citation

If you use this fine-tuned model, please cite this repository and the original Qwen3-ASR model.

@misc{visoka_qwen3_asr_17b_albanian_2026,
  author       = {Visoka, Kushtrim},
  title        = {Qwen3-ASR-1.7B Albanian},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Kushtrim/Qwen3-ASR-1.7B-Albanian},
  note         = {Fine-tuned from Qwen/Qwen3-ASR-1.7B}
}