Whisper Large V3 Turbo — Albanian

An Albanian speech recognition adaptation of OpenAI's Whisper Large V3 Turbo, developed by Kushtrim Visoka.

Evaluation

The released checkpoint recorded 9.63% validation word error rate (WER) and 0.0922 validation loss. Evaluation used an Albanian speech dataset separate from the training data. This validation dataset also guided checkpoint selection; a separate test-set result and a matched evaluation of the base model are not reported here.

Learning curves

Dots and dashed guides identify the released checkpoint. Grey segments show subsequent logging. Progress is normalized to the full run; validation curves are unsmoothed. Training loss uses a logarithmic scale, with faint raw observations and a trailing median over 1% of progress.

Vector figure · WER detail · Training-loss detail · Validation-loss detail

Intended use and limitations

The model is intended for Albanian speech transcription. Accuracy may vary with dialect, recording quality, background noise, overlapping speakers, and code-switching. Assess performance on representative audio and check transcriptions against the source recording when precise wording matters.

Attribution

Albanian adaptation and release: Kushtrim Visoka. Whisper architecture and pretrained base weights: OpenAI.

Citation

@misc{visoka_whisper_large_v3_turbo_albanian_2026,
  author       = {Visoka, Kushtrim},
  title        = {Whisper Large V3 Turbo Albanian},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Kushtrim/whisper-large-v3-turbo-sq},
  note         = {Albanian adaptation of openai/whisper-large-v3-turbo}
}