wav2vec2-large-xlsr-kurmanji
A fine-tuned Wav2Vec2 XLSR model for Automatic Speech Recognition (ASR) in Kurmanji Kurdish (kmr).
This model was fine-tuned on the Mozilla Common Voice 8.0 Kurmanji dataset using Facebook's wav2vec2-large-xlsr-53 as the base model.
Model Details
| Property | Value |
|---|---|
| Base Model | facebook/wav2vec2-large-xlsr-53 |
| Language | Kurmanji Kurdish (kmr) |
| Task | Automatic Speech Recognition |
| Dataset | Mozilla Common Voice 8.0 |
| License | Apache 2.0 |
Evaluation Results
Evaluated on the test split of Common Voice 8.0 (kmr):
| Metric | Value |
|---|---|
| WER (Word Error Rate) | 32.98% |
| CER (Character Error Rate) | 7.97% |
| Sentence Accuracy | 22.38% |
The low CER (β8%) indicates strong phoneme-level recognition. The higher WER reflects morphological variation inherent in Kurmanji (e.g., clitic boundaries, vowel alternations), which is a known challenge for low-resource Kurdish ASR.
Usage
Basic Inference
import torch
from transformers import pipeline
asr = pipeline(
task="automatic-speech-recognition",
model="salamhasan/wav2vec2-large-xlsr-kurmanji",
device=0 if torch.cuda.is_available() else -1,
)
result = asr("path/to/audio.wav")
print(result["text"])
With Chunking (Long Audio)
result = asr(
"path/to/long_audio.wav",
chunk_length_s=10,
stride_length_s=2,
)
print(result["text"])
Load Model & Processor Directly
import torch
import librosa
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
model_id = "salamhasan/wav2vec2-large-xlsr-kurmanji"
processor = Wav2Vec2Processor.from_pretrained(model_id)
model = Wav2Vec2ForCTC.from_pretrained(model_id)
model.eval()
# Load audio (must be 16kHz mono)
audio, sr = librosa.load("path/to/audio.wav", sr=16000, mono=True)
inputs = processor(
audio,
sampling_rate=16000,
return_tensors="pt",
padding=True,
)
with torch.no_grad():
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.batch_decode(predicted_ids)
print(transcription[0])
Audio Requirements
| Property | Required Value |
|---|---|
| Sample Rate | 16,000 Hz |
| Channels | Mono |
| Format | WAV / FLAC / MP3 |
Training Details
| Property | Value |
|---|---|
| Base Model | facebook/wav2vec2-large-xlsr-53 |
| Training Dataset | Mozilla Common Voice 8.0 (kmr) |
| Fine-tuning Task | CTC (Connectionist Temporal Classification) |
| Framework | π€ Transformers |
Limitations
- Optimized for read speech (Common Voice style). Performance may vary on spontaneous or conversational speech.
- Kurmanji has two main orthographic conventions (Latin and Arabic script); this model was trained on Latin script only.
- Morphological richness of Kurdish means WER is naturally higher than languages with simpler morphology.
- Low-resource language β more training data would significantly improve performance.
Language Information
Kurmanji Kurdish (kmr) is the most widely spoken dialect of the Kurdish language, spoken primarily in Northern Kurdistan (Southern Turkey), Western Kurdistan (Northern and Eastern Syria), South Kurdistan (Northern Iraq), and East Kurdistan (Western Iran) by an estimated 40β50 million speakers. It is written in a Latin-based alphabet in Northern Kurdistan and Western Kurdistan, and in an Arabic-based alphabet in South Kurdistan and East Kurdistan.
Citation
If you use this model in your research, please cite:
@misc{wav2vec2-large-xlsr-kurmanji,
author = {Salam Hasan},
title = {wav2vec2-large-xlsr-kurmanji: Fine-tuned Wav2Vec2 for Kurmanji ASR},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/salamhasan/wav2vec2-large-xlsr-kurmanji},
}
Acknowledgements
- Facebook AI Research for the XLSR-53 base model
- Mozilla Foundation for the Common Voice dataset
- Hugging Face for the Transformers library
- Downloads last month
- 16
Evaluation results
- Test WER on Common Voice 8.0test set self-reported32.980
- Test CER on Common Voice 8.0test set self-reported7.970