Itzel21/emergency-faq-speech

Emergency Room FAQ Speech

Small audio evaluation set for testing automatic speech recognition (ASR) on emergency-room FAQ questions, recorded in Taiwanese Hokkien (nan) and Mandarin (cmn). Every question is spoken in both languages, so each transcription has a matching pair of audio clips.

The questions are grouped into four categories that mirror the original hospital FAQ collection:

category	description	# questions
waiting	候診 — registration, triage, waiting	34
examination	檢查 — tests and examinations	22
observation	留觀 — observation and admission	48
other	其他 — fees, certificates, misc.	29

Structure

data/{language}/{category}/{sample_id}.wav

• language: hokkien or mandarin
• category: waiting, examination, observation, other
• sample_id: question number within its category; the Hokkien and Mandarin clips of the same question share the same sample_id

metadata.csv follows the Hugging Face AudioFolder convention (file_name column), so the dataset loads directly with:

from datasets import load_dataset
ds = load_dataset("Itzel21/emergency-faq-speech")

Columns: file_name, language, category, sample_id, transcription (original Traditional Chinese text).

Usage

Load the full set, or just one language:

from datasets import load_dataset

ds = load_dataset("Itzel21/emergency-faq-speech", split="train")
hokkien = ds.filter(lambda x: x["language"] == "hokkien")
mandarin = ds.filter(lambda x: x["language"] == "mandarin")

Each example is a dict with a decoded audio array plus its reference text:

example = ds[0]
audio = example["audio"]["array"]        # float32 waveform
sr = example["audio"]["sampling_rate"]   # 24000
reference = example["transcription"]

To benchmark an ASR system, run it over each clip and compare against transcription (e.g. with CER):

from jiwer import cer

hyp = my_asr_model(audio, sr)
score = cer(reference, hyp)

Filter by category or sample_id to reproduce results on a specific FAQ subset, or to pair up the Hokkien/Mandarin recordings of the same question (same sample_id, different language).

Dataset Creation

All audio clips are synthetic speech generated with the OpenAI Realtime API (model gpt-realtime, audio-only output as 24 kHz PCM16), prompted with language-specific instructions to read each FAQ question in Mandarin or Taiwanese Hokkien. No human speakers were recorded, so the dataset contains no personal or patient data.

Limitations

• Synthetic, not real-world speech. Clips are clean, single-speaker TTS output with no background noise, overlapping speech, or the elderly / distressed speakers typical of a real emergency room. ASR accuracy measured here is likely an upper bound on real kiosk performance.
• Hokkien quality. Hokkien was produced by a general-purpose model, not native speakers; pronunciation and word choice may be Mandarin-influenced.
• Limited speaker diversity. Clips come from synthetic voices only; results may not generalize across gender, age, or accent.

Notes

• Total: 266 clips (133 questions × 2 languages).
• Audio: WAV, 24 kHz, 16-bit, mono. Resample to 16 kHz if your ASR model expects it: ds = ds.cast_column("audio", Audio(sampling_rate=16000))
• Transcriptions are written in Traditional Chinese for both languages; Hokkien clips are labeled with the Mandarin text of the same question (semantic reference, not a Hokkien orthographic transcription).
• Intended for small-scale ASR sanity checks / regression testing, not large-scale training. Model outputs are subject to OpenAI's usage policies.

License

CC BY-NC 4.0