KazNERD_ASR

Datasets and evaluation scripts for the paper "Improved Kazakh Named Entity Transcription Using Synthetic Speech" (UBMK 2025).

[Paper]

Datasets

All datasets are hosted on Hugging Face. Transcripts are based on the KazNERD corpus (Kazakh television news text rich in named entities). sentence_id corresponds to the KazNERD sentence ID without its trailing AID (e.g. AID343740 ↔ AID343740AID).

Dataset	Type	Splits	Utterances	Size
issai/KazNERD_H	Human-recorded speech	sample1–sample6	531	0.1 GB
issai/kaznerd1	Synthetic speech	train / valid / test	39,493 / 4,908 / 4,385	10.8 GB
issai/kaznerd2	Synthetic speech	train	35,609	7.7 GB

• KazNERD_H: KazNERD test sentences read aloud by human speakers, used as the real-speech test set.
• kaznerd1: synthetic speech of the original KazNERD sentences; its splits follow the KazNERD train/validation/test splits.
• kaznerd2: synthetic speech of modified KazNERD training sentences (the transcripts differ from the original KazNERD text), used as additional training data.

Synthetic audio was generated with Kazakh TTS models; the model and speaker for each utterance are given in the tts_model and speaker_id fields.

Audio is 16 kHz mono WAV, stored in Parquet. Fields:

• audio: audio (WAV)
• text: Kazakh transcript
• sentence_id: KazNERD sentence ID
• speaker_id, gender, tts_model: synthetic datasets only

Loading

from datasets import load_dataset

human = load_dataset("issai/KazNERD_H")            # splits: sample1 ... sample6
kaznerd1 = load_dataset("issai/kaznerd1")          # splits: train, valid, test
kaznerd2 = load_dataset("issai/kaznerd2", split="train")

sample = kaznerd1["train"][0]
print(sample["text"], sample["tts_model"], sample["speaker_id"])

Evaluation scripts

The test/ folder contains the scripts used to evaluate fine-tuned Whisper checkpoints. Each script transcribes a test set and reports WER in two settings: normalized (lowercased, punctuation removed) and un-normalized (casing and basic punctuation kept).

Script	Test set
test/human_test.py	KazNERD_H (human-recorded speech)
test/kaznerd_test.py	Synthetic KazNERD test set (kaznerd1 test split)
test/ksc_test.py	KSC2 crowdsourced test set
test/fleurs_test.py	FLEURS Kazakh (kk_kz) test set
test/find_cer.py	CER over a folder of prediction files

Checkpoint and data paths at the bottom of each script are set to the original experiment environment; change them to your local paths before running.

Citation

@inproceedings{meiramov2025kaznerdasr,
  author    = {Meiramov, Rakhat and Varol, Huseyin Atakan},
  title     = {Improved Kazakh Named Entity Transcription Using Synthetic Speech},
  booktitle = {2025 10th International Conference on Computer Science and Engineering (UBMK)},
  pages     = {1629--1633},
  year      = {2025},
  doi       = {10.1109/UBMK67458.2025.11242232}
}

Please also cite KazNERD, the source of the transcripts.