LocalDoc/azerbaijani_asr

Azerbaijani ASR Dataset

Dataset Description

This dataset contains Azerbaijani speech data for Automatic Speech Recognition (ASR) tasks.

Dataset Summary

• Language: Azerbaijani (az)
• Task: Automatic Speech Recognition
• Total Duration: ~328 hours
• Total Samples: ~345,643 audio-text pairs
• Audio Format: WAV, 16kHz sampling rate
• License: CC-BY-4.0

Dataset Structure

Each audio segment is specially numbered so that you can merge them if you need to create longer segments.

DatasetDict({
    train: Dataset({
        features: ['audio', 'text', 'duration', 'audio_file'],
        num_rows: [total_samples]
    })
})

Features

• audio: Audio file (WAV format, 16kHz)
• text: Transcription of the audio (Azerbaijani text)
• duration: Duration of audio clip in seconds
• audio_file: Original filename of the audio segment
• text_norm: The same sentence in spoken form — digits, dates, times, percents and symbols written out as they are pronounced
• norm_status: How text_norm was produced for this row (see below)
• norm_replacements: JSON list of the fragments that were rewritten, each with the original, the reading used and other likely readings (alternatives)

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("LocalDoc/azerbaijani_asr")

# Access train split
train_dataset = dataset["train"]

# Example usage
for example in train_dataset:
    audio = example["audio"]
    text = example["text"]
    duration = example["duration"]

Duration Distribution

• 0-2 sec: 36.1%
• 2-5 sec: 47.2%
• 5-10 sec: 14.6%
• 10-20 sec: 2.0%
• 20+ sec: 0.1%

Use Cases

• Fine-tuning Whisper models for Azerbaijani
• Training new ASR models for Azerbaijani language
• Benchmarking ASR performance on Azerbaijani

Spoken-form Normalization

The text column is unchanged: it is the original transcription, written the way text is written. Alongside it, text_norm gives the same sentence the way it is actually said out loud, so that an ASR model is trained on the words that are in the audio rather than on their written shorthand.

Rows	345,643
Hours	329.6
Usable for spoken-form training (text_norm not null)	345,629 rows, 329.6 h
Rows with alternatives	698

What changes, with examples from this dataset:

• Percents are spelled out: 3% -> üç faiz.
• Times are read the way they are said: 9:00 -> doqquz.
• A long digit string is read in dictated groups, separated by commas, the way a phone number or a code is actually said: 182006026 -> yüz səksən iki, sıfır sıfır altı, sıfır iyirmi altı.
• A bare thousand is min, never bir min: 10.000 -> on min.
• Decimals are read as values, not as prices: 4.0 -> dörd tam sıfır.
• A hyphen stays between the two numbers of a range: 2-3 -> iki-üç.
• Numbers become words, and an attached suffix is merged in with the right vowel harmony: 3-ü -> üçü.
• Zero is always sıfır, and Latin names and loanwords stay exactly as they are written.
• Brand names are not merged into one spelling. In a call-centre corpus bir bank usually means the bank called Birbank; in read speech it almost always means «some bank», so nothing is unified here.

Every reading is checked automatically: the number words are parsed back into digits and compared with the source, and a reading that loses or invents a number is rejected and re-asked. Whatever still fails that check is marked failed and carries text_norm = null (7 of the rows here).

norm_status per row:

status	rows	hours	meaning
rules_only	332,434	307.7	no digits or symbols to rewrite — the text was already spoken form
ok	13,195	21.9	rewritten by the model and validated
failed	7	0.0	no reading passed validation; text_norm is null
skipped_mixed	7	0.0	the text mixes scripts (Cyrillic), left out

Where the spoken form is genuinely uncertain — a version number, a date, a long code — norm_replacements carries the other likely readings in alternatives, so they can be re-ranked against the audio instead of being guessed from text.

The normalization follows the same written-to-spoken standard as our Azerbaijani call-centre and YouTube corpora — same rules, same validator, same prompt except for the two sentences that name the domain — so the datasets can be mixed in one training run.

Citation

If you use this dataset, please cite:

@dataset{azerbaijani_asr_2025,
  title={Azerbaijani ASR Dataset},
  author={LocalDoc},
  year={2025},
  url={https://huggingface.co/datasets/LocalDoc/azerbaijani_asr}
}