ddehun/korean-weather-asr

Speech Recognition Benchmark for Korean Meteorological Experts

This repository provides the dataset accompanying the paper:
Evaluating Automatic Speech Recognition Systems for Korean Meteorological Experts, EMNLP2025 (short, findings).

Overview

This dataset is created to evaluate automatic speech recognition (ASR) systems in the domain of Korean meteorology. Unlike general-purpose speech benchmarks, it emphasizes:

• Frequent use of specialized meteorological terminology
• Korean linguistic challenges such as spacing and agglutinative morphology
• Utterances reflecting realistic expert queries to weather information systems

The dataset enables reliable benchmarking of ASR models for weather-related applications.

Structure

• Audio Data: WAV files containing meteorological queries (commonly sampled at 48kHz, though sampling rate may vary by file).
• Transcriptions: Human-verified transcriptions aligned with the audio.

Split

• test – currently provided only for evaluation.
• train – will be released soon to support model development and fine-tuning.

Example Data

Each entry in the dataset follows the structure below:

{
  "index": 0,
  "audio_fname": "0.wav",
  "text": "8년 전 3월 이전 하순 통영에서 지점별 가장 1시간최다강수량이 많은 날 검색해주세요.",
  "sr": 48000,
  "duration": 6.49
}

• index: unique ID
• audio_fname: corresponding audio file name (WAV format)
• text: ground-truth transcription
• sr: sampling rate (Hz), typically 48kHz but not guaranteed
• duration: audio length in seconds

Statistics

Metric	Value
Num. samples	5,492
Utterance time (sec)	7.05 ± 2.55
└ min / max time	0.92 / 29.98
Avg. chars / words	24.49 ± 15.62 / 7.59 ± 4.34
Unique words	4,955
Absent ratio (%)	24.86

Citation

@article{park2024evaluating,
  title={Evaluating Automatic Speech Recognition Systems for Korean Meteorological Experts},
  author={Park, ChaeHun and Cho, Hojun and Choo, Jaegul},
  journal={arXiv preprint arXiv:2410.18444},
  year={2024}
}