Quyashbek/wav2vec2-xls-r-300m-karakalpak

🎙️ Wav2Vec2 XLS-R 300M Karakalpak ASR

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m for Automatic Speech Recognition (ASR) in the Karakalpak language.

---

👤 Founder

Quyashbek Allanazarov

---

🤝 Acknowledgements

• 🎓 New Uzbekistan University — for knowledge, research environment, and academic direction
• 🏦 Xalq Banki AI Lab — for providing GPU resources and supporting the technical direction of the project

---

📊 Model Performance

Evaluation was performed on a held-out test set.

Metric	Score
WER (Word Error Rate)	21.21%
CER (Character Error Rate)	4.34%

Test Details

• Total samples: 504
• Successful transcriptions: 504
• Missing files: 0
• Errors: 0
• Sampling rate: 16 kHz

---

⚙️ Model Details

• Base model: facebook/wav2vec2-xls-r-300m
• Architecture: Wav2Vec2ForCTC
• Language: Karakalpak
• Task: Speech-to-Text (ASR)
• Framework: PyTorch + Hugging Face Transformers

---

🧪 Inference Example

import torch
import librosa
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC

model_id = "Quyashbek/wav2vec2-xls-r-300m-karakalpak-asr"

processor = Wav2Vec2Processor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)

speech, sr = librosa.load("your_audio.wav", sr=16000)

inputs = processor(speech, sampling_rate=16000, return_tensors="pt", padding=True)

with torch.no_grad():
    logits = model(inputs.input_values).logits

pred_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(pred_ids)[0]

print(transcription)