Speech › Synthetic K-OR Speech Audio › k-or-speech-asr@0

Korean operating-room speech recognition preview

Data: khdp.net/database/data-search-detail/774/Synthetic-K-OR-Speech-Audio/1.0.0 · Apache-2.0 with a use restriction · asr · primary metric: cer · spec 27b75c766635

Each of the 8,000 clips (2,000 Korean operating-room utterances read by 4 synthetic voices) is transcribed once with language set to Korean. The score is character error rate (CER, lower is better) after the same normalisation of reference and transcript: punctuation removed, numbers read out in Hangul, spaces ignored; word error rate is also reported. 95% intervals come from a bootstrap over utterances, with the 4 voices of an utterance kept together. Results by voice, code-switching style, intent and speaker role are on each result page.

Synthetic speech has no room noise, reverberation, overlapping talk or real urgency, so error rates here are lower than they would be in real operating rooms.

Use restriction (dataset license): the audio may be used for speech recognition research, benchmarking, evaluation and zero-shot testing. It must not be used to train, fine-tune or develop text-to-speech, voice-cloning or speaker-identification systems, nor to create any derivative voice asset. Redistribution must keep this restriction.

Data

Synthetic K-OR Speech Audio (speech). License: Apache-2.0 with a use restriction. Source: https://khdp.net/database/data-search-detail/774/Synthetic-K-OR-Speech-Audio/1.0.0

Seoul National University Hospital, Department of Anesthesiology and Pain Medicine. Synthetic K-OR Speech Audio: A Synthetic Korean Operating Room Speech Audio Dataset (version 1.0.0). KHDP. 2026. doi:10.23421/2026.0774

Task specification

{
  "dataset": "k-or-speech",
  "instance_id": "<utterance id>-<voice>, e.g. MMO-M00002-chul",
  "kind": "asr",
  "license": "Audio Apache-2.0 with a use restriction: it may be used for ASR/STT research, benchmarking, evaluation and zero-shot testing; it must not be used to train, fine-tune or develop text-to-speech, voice-cloning or speaker-identification systems, nor to create any derivative voice asset; redistribution must keep this restriction. Text metadata CC BY 4.0 (MM-OR-derived rows Apache-2.0).",
  "limitations": [
    "Synthetic speech (Qwen3-TTS) has no room noise, reverberation, overlap or real urgency, so error rates here understate those in real operating rooms.",
    "One voice (chul) is a voice clone of an SNUH anesthesiologist; only ethan and maia carry intent-conditioned prosody.",
    "The audio was made with Qwen3-TTS (Alibaba); an ASR model from the same developer may be favoured.",
    "Unit spellings vary inside the references themselves (cc, 시씨, 씨씨); they are scored as written."
  ],
  "metrics": {
    "bootstrap": {
      "by": "utterance (the 4 voices of one utterance are resampled together)",
      "n": 1000,
      "seed": 0
    },
    "breakdowns": [
      "voice",
      "cs_style",
      "intent",
      "speaker",
      "digits in reference"
    ],
    "cer": "sum of character edit distances / sum of reference characters over all clips, spaces removed",
    "missing": "a clip without a transcript (API error) counts as an empty transcript, so every reference character is a deletion",
    "primary": "cer",
    "ser": "share of clips with at least one character error",
    "wer": "sum of word edit distances / sum of reference words over all clips (whitespace tokens; sensitive to Korean spacing)"
  },
  "normalisation": {
    "applied_to": "reference and transcript alike",
    "steps": [
      "Unicode NFKC (NFC plus compatibility folding of full-width letters, digits and unit symbols), then lowercase",
      "remove bracketed tags [..], <..>, (..)",
      "numbers: commas inside numbers dropped; Arabic numerals read in Hangul, Sino-Korean (5 번 -> 오 번, 3만 -> 삼만), native before 번째, 시간, 개, 명, 장, 살, 마리, 가지, 군데 (12개 -> 열두개); decimals digit by digit with the spoken point optional, because two of the four voices say 점 and two mostly do not (7.5, 칠 점 오 and 칠오 all become 칠오); digits attached to Latin letters (SpO2, EtCO2, R2P3) stay as written; % -> 퍼센트",
      "every punctuation, symbol and mark character becomes a space; whitespace collapses",
      "English code-switched terms stay as written (lowercased): Hangul for a Latin reference term (비피 for BP) or the reverse counts as an error"
    ],
    "version": "kor-norm-0"
  },
  "preview": true,
  "primary_metric": "cer",
  "protocol_source": "The dataset's usage notes (KHDP Synthetic-K-OR-Speech-Audio 1.0.0) recommend CER and WER against the v1.0 text, reported by intent, code-switching style, voice and speaker role. They define no text normalisation; the rules below are medbench.kr's.",
  "request": {
    "audio": "the release WAV as is (24 kHz, 16-bit, mono)",
    "endpoint": "OpenAI-compatible POST /audio/transcriptions, one clip per request",
    "language": "ko",
    "response_format": "json",
    "temperature": 0
  },
  "slug": "k-or-speech-asr",
  "source": {
    "archive": "synthetic_kor_speech_audio_v1.zip",
    "khdp_code": "Synthetic-K-OR-Speech-Audio",
    "reference": "metadata.csv column text (identical to Synthetic-K-OR-Speech-Corpus 1.0.0 for all 2,000 ids, checked 2026-10-29)",
    "sha256": "fb60266a06cf710535d3e4468b8a1748adda6b58de231e05a5e407b5c2de772d",
    "version": "1.0.0"
  },
  "split": "test-only (all 8,000 clips = 2,000 utterances x 4 synthetic voices; fixed benchmark, nothing is trained)",
  "title": "Korean operating-room speech recognition",
  "version": 0
}

JSON: /v1/tasks/k-or-speech-asr · /v1/leaderboard?task=k-or-speech-asr@0