TonyFANgr/localaiiaasr

Emergency Room FAQ ASR Evaluation Set

Small audio evaluation set for testing automatic speech recognition (ASR) on emergency-room FAQ questions, recorded in Taiwanese Hokkien (nan), Mandarin (cmn), and code-switched Hokkien/Mandarin (mixed). Most questions are spoken in both Hokkien and Mandarin, so most transcriptions have a matching pair of audio clips; a small subset also has a mixed-language clip.

The questions come from eval/FAQ_dataset in the local_aiia project, grouped into four categories that mirror the original hospital FAQ collection:

category	source file	# questions
waiting	1waiting.json	34
examination	2examination.json	22
observation	3observation.json	48
other	4other.json	29

Structure

{language}/{category}/id_{id}_{english_slug}.wav

• language: Hokkien, Mandarin, or Mixed
• category: 1waiting, 2examination, 3observation, 4other
• id: matches the id field in the corresponding eval/FAQ_dataset/*.json file
• english_slug: short English gist of the question, for readability only — the authoritative transcription is in metadata.csv

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("TonyFANgr/localaiiaasr")

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

Usage

Load the full set, or just one language:

from datasets import load_dataset

ds = load_dataset("TonyFANgr/localaiiaasr", split="train")
hokkien = ds.filter(lambda x: x["language"] == "hokkien")
mandarin = ds.filter(lambda x: x["language"] == "mandarin")
mixed = ds.filter(lambda x: x["language"] == "mixed")

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"]   # 16000
reference = example["transcription"]

To benchmark an ASR system, run it over each clip and compare against transcription (e.g. with CER / semantic similarity, as in eval/asr/metrics.py):

from jiwer import cer

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

Filter by category (waiting, examination, observation, other) 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).

Notes

• Total: 274 clips (133 questions × 2 languages, plus 8 mixed-language clips).
• Intended for small-scale ASR sanity checks / regression testing, not large-scale training.