Kartal-Ol/ASR-AZB

South Azerbaijani ASR Models

Model checkpoints for the INTERSPEECH 2026 paper “Preserving the Iranian Turkic Language: Community-Driven ASR Datasets and Benchmarking for South Azerbaijani.”

Training, evaluation, normalization, and reproducibility code is available at Kartalol/Kartalol-azb-asr.

Available checkpoints

Folder	Model	Training setup / initialization
whisper-tiny	Whisper Tiny	Community/book data
whisper-base	Whisper Base	Community/book data
whisper-base-full	Whisper Base	Full dataset; preserved from the original repository root
whisper-Small	Whisper Small	Community/book data
whisper-Small-Farsi	Whisper Small	Farsi cross-lingual initialization
whisper-small-north-azerbaijani	Whisper Small	North Azerbaijani cross-lingual initialization
whisper-small-turkish	Whisper Small	Turkish cross-lingual initialization
whisper-Small-Arabic	Whisper Small	Arabic cross-lingual initialization

The original root-level Full-dataset Whisper Base files remain available for backwards compatibility. whisper-base-full is their organized mirror.

Loading a checkpoint

import librosa
import torch
from transformers import AutoProcessor, WhisperForConditionalGeneration

repo_id = "Kartal-Ol/ASR-AZB"
subfolder = "whisper-base-full"  # choose a folder from the table above

processor = AutoProcessor.from_pretrained(repo_id, subfolder=subfolder)
model = WhisperForConditionalGeneration.from_pretrained(
    repo_id,
    subfolder=subfolder,
)

audio, _ = librosa.load("audio.ogg", sr=16_000, mono=True)
inputs = processor(audio, sampling_rate=16_000, return_tensors="pt")

device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device).eval()
with torch.inference_mode():
    predicted_ids = model.generate(inputs.input_features.to(device))

print(processor.batch_decode(predicted_ids, skip_special_tokens=True)[0])

Evaluation

The paper reports Word Error Rate (WER) and Character Error Rate (CER) as percentages. Deletion/Insertion Ratio (DIR) is reported as a raw ratio. Refer to the GitHub repository for the official benchmark tables and evaluation command.

Data

• Kartal-Ol/azb-asr-corpus
• yoyo-research-group/south-azerbaijani-asr
• Kartal-Ol/AZB-ASR-Gold-Testset

The GoldSet is held out for evaluation and must not be included in training.

Citation

@inproceedings{farsi2026preserving,
  title     = {Preserving the Iranian Turkic Language: Community-Driven ASR
               Datasets and Benchmarking for South Azerbaijani},
  author    = {Farsi, Farhan and Bali, Shayan and Nourmohammadi Khiarak, Jalil
               and Aref, Mohammad Hossein and Akbari Saeed, Taher},
  booktitle = {Proceedings of INTERSPEECH 2026},
  year      = {2026}
}

We thank the Kartal Ol Foundation and the community contributors who made these resources possible.