glossKit-ASR/wav2vec2-large-xlsr-53-zza

Model Description

This model is a fine-tuned version of ShuanOsmanKarim/mohammadirad-wav2vec2-haw for automatic speech recognition, trained primarily on speech labeled with ISO 639-3 zza (ZZA).

Model Performance

• Word Error Rate (WER): 0.9373 (93.73%)
• Character Error Rate (CER): 0.4099 (40.99%)
• Test Samples: 67

Training Details

• Base Model: ShuanOsmanKarim/mohammadirad-wav2vec2-haw
• Language: ZZA (zza)
• Fine-tuning Framework: PyTorch / HuggingFace Transformers

Data provenance

Training segments were contributed through the following GlossKit projects: Zazakî Documentation Project. They are aggregated under the training language code zza.

Languages in the training set

Language labels come from ACTIVE contributing GlossKit projects; minutes are summed from metadata rows that include projectId (no estimates or splits):

• Zazakî (23min)

Contributing consultants

Consultant IDs (project aliases), age, and gender only — personal names are not listed:

• GM_f2 (48F)
• M004 (0F)
• S002 (60F)
• T003 (30M)

Contributing users

People who own or collaborate on the contributing GlossKit projects (alphabetical by display name):

• Mahîr Dogan
• Shuan Karim

Usage

from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import torch

processor = Wav2Vec2Processor.from_pretrained("glossKit-ASR/wav2vec2-large-xlsr-53-zza")
model = Wav2Vec2ForCTC.from_pretrained("glossKit-ASR/wav2vec2-large-xlsr-53-zza")

# Process audio and transcribe
inputs = processor(audio, sampling_rate=16_000, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.decode(predicted_ids[0])

Citation

If you use this model in your research, please cite:

@misc{glosskit-asr-zza,
  title={GlossKit ASR Model (ZZA)},
  author={Mahîr Dogan and Shuan Osman Karim},
  year={2026},
  url={https://huggingface.co/glossKit-ASR/wav2vec2-large-xlsr-53-zza}
}

License

MIT License