Qwen3-ASR-0.6B Albanian

This is a fine-tuned variant of Qwen/Qwen3-ASR-0.6B for Albanian (sq / Shqip) automatic speech recognition.

It is trained on a couple thousand well-curated hours of Albanian audio and transcripts.

Intended use

• ASR / transcription for Albanian speech (general-purpose).
• Works best on clean speech audio (mono, 16 kHz) with transcripts matching the dataset conventions.

Limitations

• Performance depends heavily on domain/accent/noise conditions; evaluate on your target audio before deployment.
• Streaming / timestamps support is provided by the upstream toolkit; quality of timestamps depends on the forced-aligner setup and audio conditions.

Base model (Qwen3-ASR)

From the upstream model card, Qwen3-ASR provides:

• Language identification + ASR in one model (supports many languages/dialects).
• Offline + streaming inference modes via the qwen-asr toolkit.
• Optional timestamps via the Qwen forced-aligner.

See the full upstream documentation here:

• https://huggingface.co/Qwen/Qwen3-ASR-0.6B

How to use (Transformers backend via qwen-asr)

Install:

pip install -U qwen-asr

Load this model from a local folder (this directory) and transcribe:

import torch
from qwen_asr import Qwen3ASRModel

model = Qwen3ASRModel.from_pretrained(
    "./",
    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",  # or None for auto language ID
)

print(results[0].language)
print(results[0].text)

Training data

• Language: Albanian (sq)
• Duration: A couple thousand well-curated hours
• Type: supervised ASR fine-tuning (audio → transcript)

Training recipe (high level)

Extracted from saved transformers training artifacts in this folder:

• Epochs: 2
• Learning rate: 2e-5
• Batch size (per device): 8
• Grad accumulation: 1
• Precision: BF16
• Optimizer: adamw_torch_fused
• Seed: 42

Notes

• Qwen3-ASR fine-tuning data typically uses transcripts with a language prefix like language Albanian<asr_text>... (see upstream docs); this fine-tune follows that convention.

Development and attribution

• Original architecture and base weights: Qwen Team, through Qwen/Qwen3-ASR-0.6B
• Albanian adaptation, fine-tuning, training-data curation, 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_06b_albanian_2026,
  author       = {Visoka, Kushtrim},
  title        = {Qwen3-ASR-0.6B Albanian},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Kushtrim/Qwen3-ASR-0.6B-Albanian},
  note         = {Fine-tuned from Qwen/Qwen3-ASR-0.6B}
}