version string | total_terms int64 | last_updated timestamp[s] | description string | terms list | changelog list |
|---|---|---|---|---|---|
2026.05 AI-Optimized Full v6 | 238 | 2026-05-08T00:00:00 | "保育用語集 AI最適化版 v6。全238項目に stt_oral_patterns(口語表現パターン)(...TRUNCATED) | [{"term":"愛着","reading":"あいちゃく","category":"発達・心理","priority":"高","tags":[(...TRUNCATED) | [{"version":"v6","date":"2026-05-08T00:00:00","changes":["stt_oral_patterns を全238項目に拡張(...TRUNCATED) |
保育用語STT最適化データセット
Japanese Childcare Vocabulary Dataset for STT Optimization
🇯🇵 日本語
概要
このデータセットは、日本の保育現場における音声認識(STT)の精度向上を目的として構築された、保育専用の構造化語彙集です。
一般的なSTTエンジン(Google・Amazon・OpenAI Whisper等)は、保育現場で日常的に使われる専門用語を正確に認識できないという問題があります。たとえば:
| 実際の発話 | STTが誤認識する例 |
|---|---|
| 誤飲(ごいん) | 語韻・誤音 |
| 午睡(ごすい) | 誤推・誤睡 |
| 降園(こうえん) | 公演・後援 |
| 点呼(てんこ) | 転校・天候 |
| 視診(ししん) | 試診・指診 |
このデータセットは、こうした保育特有の同音異義語・誤変換パターンを体系的に収録し、STTエンジンのカスタム辞書・RAG(検索拡張生成)・AIコール(音声電話応答)に活用できるよう設計されています。
背景・開発経緯
開発元である株式会社インフォマートは、保育施設向けヒヤリ・ハット報告AIアプリ「ヒヤリん」を運営しています。ヒヤリんはSTT(音声入力)をユーザーインターフェースとして採用しており、保育士が音声で事故ヒヤリ・ハットを報告できるシステムです。
実際の運用の中で「保育専用のSTT辞書が存在しない」という課題に直面し、このデータセットを独自に構築しました。
また、2023年のこども家庭庁ガイドライン施行(不適切保育の防止・自己点検の義務化)を受け、不適切保育・権利擁護カテゴリを重点的に収録しています。
データセット仕様
| 項目 | 内容 |
|---|---|
| 総用語数 | 238語 |
| バージョン | 2026.05 AI-Optimized Full v6 |
| 形式 | JSON |
| 言語 | 日本語 |
| ライセンス | CC BY-NC 4.0 |
カテゴリ分布
| カテゴリ | 件数 | 説明 |
|---|---|---|
| 安全管理 | 38 | ヒヤリ・ハット・置き去り・誤飲・窒息等 |
| 健康・医療 | 35 | 感染症・緊急対応・日常症状等 |
| 発達・心理 | 29 | 発達段階・障害・心理的支援等 |
| 不適切保育・権利擁護 | 19 | 体罰・暴言・自己点検・実地指導等 |
| 制度・法令 | 17 | 児童福祉法・こども基本法・認可制度等 |
| 保育形態 | 16 | 延長保育・一時保育・異年齢保育等 |
| 食育・栄養 | 14 | 離乳食・アレルギー除去食・代替食等 |
| 日課・活動 | 14 | 午睡・散歩・お集まり等 |
| 日常症状・口語 | 11 | 発熱・嘔吐・下痢・欠席連絡等 |
| 欠席・登降園連絡 | 7 | お迎え時間変更・代理お迎え等 |
| その他 | 38 | 保護者支援・職員・施設・記録等 |
優先度分布
| 優先度 | 件数 | 基準 |
|---|---|---|
| 最高 | 53 | 生命・安全・法令直結。STT誤認識で重大リスク |
| 高 | 80 | 現場での頻出度・重要度が高い用語 |
| 中 | 83 | 標準的な保育用語 |
| 低 | 22 | 補足的・参照用の用語 |
フィールド仕様
各用語は以下のフィールドで構成されています。
{
"term": "用語(表記)",
"reading": "よみがな(ひらがな)",
"category": "カテゴリ",
"priority": "優先度(最高/高/中/低)",
"tags": ["タグ1", "タグ2"],
"description": "用語の説明文",
"mis_conversions": [
"STTが誤認識しやすいパターン1(同音異義語)",
"STTが誤認識しやすいパターン2(近似音)",
"STTが誤認識しやすいパターン3(語尾変化)"
],
"stt_oral_patterns": [
"保護者・保育士が実際に発話するパターン1",
"保護者・保育士が実際に発話するパターン2"
],
"contexts": [
"実際の保育現場での使用文例1",
"実際の保育現場での使用文例2"
],
"related_terms": ["関連用語1", "関連用語2"]
}
mis_conversionsの設計思想
3パターンで構成されています:
- [0] 同音異義語:全く別の意味の言葉に化ける(最重要)
- [1] 近似音・音韻ずれ:子音・母音が1つずれる
- [2] 口語短縮・語尾変化:STTが語尾を切る・変形する
stt_oral_patternsの設計思想
電話口での実際の発話は文章として完結しないことが多いため、以下のバリエーションを収録:
- 体言止め(「熱がある」)
- 過去形(「熱が出た」)
- 形容詞表現(「熱っぽい」)
- 数値を含む表現(「37度5分あります」)
- 文節途中で終わる電話特有の話し方(「熱が下がらなくて」)
使用例
1. STTカスタム辞書への組み込み
import json
with open("hoiku_yougo_v6.json", encoding="utf-8") as f:
dataset = json.load(f)
# 最高優先度の用語を抽出してカスタム辞書を構築
high_priority_terms = [
{
"term": t["term"],
"reading": t["reading"],
"mis_conversions": t["mis_conversions"]
}
for t in dataset["terms"]
if t["priority"] == "最高"
]
print(f"最高優先度用語数: {len(high_priority_terms)}")
# → 53件
2. RAG(検索拡張生成)への活用
# 安全管理カテゴリの用語でRAG用ベクトルDBを構築
safety_terms = [
t for t in dataset["terms"]
if t["category"] == "安全管理"
]
for term in safety_terms:
# contextsをRAGの検索対象テキストとして使用
for context in term["contexts"]:
vector_db.add(
text=context,
metadata={"term": term["term"], "priority": term["priority"]}
)
3. AI電話応答(AIコール)での活用
# 保護者の発話パターンを使って意図分類
oral_patterns = {}
for term in dataset["terms"]:
for pattern in term.get("stt_oral_patterns", []):
oral_patterns[pattern] = {
"term": term["term"],
"category": term["category"],
"priority": term["priority"]
}
# 例:「今日お休みします」→ 欠席連絡カテゴリに分類
データサンプル
{
"term": "誤飲",
"reading": "ごいん",
"category": "安全管理",
"priority": "最高",
"tags": ["事故防止", "乳幼児", "異物"],
"description": "様々な物への興味から口に入れ誤って飲み込むこと。乳幼児は特に注意が必要で、重大事故につながる場合がある。",
"mis_conversions": ["語韻", "誤音", "五飲"],
"stt_oral_patterns": [
"誤飲しました",
"飲み込んでしまって",
"何か口に入れて",
"異物を飲んで"
],
"contexts": [
"「誤飲を防ぐため、小さな部品は乳幼児の手の届かない場所に保管してください」",
"「誤飲事故が発生した場合は直ちに救急対応を取り、飲み込んだものを確認します」",
"「誤飲はヒヤリ・ハット報告で最も多いカテゴリのひとつです」"
],
"related_terms": ["ヒヤリ・ハット", "窒息", "アナフィラキシー"]
}
ライセンス
CC BY-NC 4.0(クリエイティブ・コモンズ 表示-非営利 4.0)
- ✅ 研究・学術目的での利用:自由
- ✅ 非営利のオープンソースプロジェクト:自由
- ✅ クレジット表示のうえでの改変・再配布:自由
- ❌ 商業製品・商業サービスへの組み込み:要相談
商用利用をご検討の場合は、下記連絡先までお問い合わせください。
引用
本データセットを研究・論文に使用する場合は、以下の形式でご引用ください:
@dataset{infomart_hoiku_stt_2026,
title = {保育用語STT最適化データセット / Japanese Childcare Vocabulary Dataset for STT},
author = {株式会社インフォマート / Infomart Co., Ltd.},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/IDEMITSU/hoiku-yougo-stt-ja},
note = {ヒヤリん(Hiyarin)開発知見より構築}
}
お問い合わせ・商用利用・連携
- Web: https://hiyarin.com
- Email: support@hiyarin.com
- 運営: 株式会社インフォマート(東京都千代田区麹町4-2-2 麹町ビル4F)
保育AIの共同研究・データセット拡充への参加・商用ライセンス提供についてもお気軽にご連絡ください。
🇺🇸 English
Overview
This dataset is a structured vocabulary resource for Japanese childcare (保育/Hoiku) settings, specifically optimized for Speech-to-Text (STT) accuracy improvement.
General-purpose STT engines (Google, Amazon, OpenAI Whisper, etc.) frequently misrecognize domain-specific terminology used in Japanese childcare environments. For example:
| Intended Term | Common STT Misrecognition |
|---|---|
| 誤飲 (goIn) — accidental ingestion | 語韻 (goIn) — phonetic tone |
| 午睡 (goSui) — nap time | 誤推 (goSui) — erroneous inference |
| 降園 (koUen) — leaving nursery | 公演 (koUen) — public performance |
| 点呼 (Tenko) — roll call | 転校 (Tenko) — school transfer |
| 視診 (shiShin) — visual inspection | 試診 (shiShin) — trial examination |
This dataset systematically captures childcare-specific homophones and misrecognition patterns in Japanese, designed for integration into STT custom dictionaries, RAG (Retrieval-Augmented Generation) systems, and AI phone response (voice call automation) applications.
Background
The dataset was developed by Infomart Co., Ltd., operators of Hiyarin — an AI-powered near-miss (Hiyari-Hatto / ヒヤリ・ハット) incident reporting application for Japanese childcare facilities. Hiyarin uses STT as its primary user interface, enabling nursery teachers to report safety incidents by voice.
During real-world operation, we identified a critical gap: no Japanese STT vocabulary resource existed for the childcare domain. This dataset was built from that operational knowledge.
In response to the 2023 Japanese Cabinet Office (こども家庭庁 / Children and Families Agency) guidelines mandating childcare facility self-inspections for inappropriate care prevention, we also added a dedicated category for inappropriate care and child rights protection terminology.
Dataset Specifications
| Item | Detail |
|---|---|
| Total entries | 238 terms |
| Version | 2026.05 AI-Optimized Full v6 |
| Format | JSON |
| Language | Japanese (ja) |
| License | CC BY-NC 4.0 |
Category Distribution
| Category (Japanese) | Category (English) | Count |
|---|---|---|
| 安全管理 | Safety Management | 38 |
| 健康・医療 | Health & Medical | 35 |
| 発達・心理 | Child Development & Psychology | 29 |
| 不適切保育・権利擁護 | Inappropriate Care & Child Rights | 19 |
| 制度・法令 | Laws & Regulations | 17 |
| 保育形態 | Childcare Formats | 16 |
| 食育・栄養 | Food Education & Nutrition | 14 |
| 日課・活動 | Daily Activities | 14 |
| 日常症状・口語 | Daily Symptoms & Colloquial | 11 |
| 欠席・登降園連絡 | Absence & Pick-up Notifications | 7 |
| Others | Parental Support, Facilities, etc. | 38 |
Priority Distribution
| Priority | Count | Criteria |
|---|---|---|
| 最高 (Critical) | 53 | Life safety / legal compliance. Misrecognition poses serious risk |
| 高 (High) | 80 | High frequency or importance in daily childcare operations |
| 中 (Medium) | 83 | Standard childcare terminology |
| 低 (Low) | 22 | Supplementary or reference terms |
Field Schema
Each entry contains the following fields:
{
"term": "Term (written form)",
"reading": "Phonetic reading (hiragana)",
"category": "Category",
"priority": "Priority level (最高/高/中/低)",
"tags": ["tag1", "tag2"],
"description": "Definition and contextual explanation",
"mis_conversions": [
"Common STT misrecognition pattern 1 (homophone)",
"Common STT misrecognition pattern 2 (near-phonetic)",
"Common STT misrecognition pattern 3 (truncation/variation)"
],
"stt_oral_patterns": [
"Actual spoken pattern by parents or nursery staff 1",
"Actual spoken pattern by parents or nursery staff 2"
],
"contexts": [
"Real-world usage example in childcare settings 1",
"Real-world usage example in childcare settings 2"
],
"related_terms": ["related term 1", "related term 2"]
}
Design Notes: mis_conversions
Each entry includes 3 misrecognition patterns:
- [0] Homophones: Words that sound identical but have completely different meanings
- [1] Near-phonetic variants: One phoneme shifted (consonant or vowel substitution)
- [2] Colloquial truncations: How STT engines cut or alter word endings in natural speech
Design Notes: stt_oral_patterns
Phone conversations rarely follow complete sentence structures. Patterns include:
- Noun-phrase endings ("熱がある" = "has a fever")
- Past-tense forms ("熱が出た" = "fever came on")
- Adjectival expressions ("熱っぽい" = "feels feverish")
- Expressions with numeric values ("37度5分あります" = "37.5°C")
- Mid-sentence phone speech patterns ("熱が下がらなくて" = "the fever won't go down...")
Usage Examples
1. Building a Custom STT Dictionary
import json
with open("hoiku_yougo_v6.json", encoding="utf-8") as f:
dataset = json.load(f)
# Extract critical-priority terms for custom dictionary
critical_terms = [
{
"term": t["term"],
"reading": t["reading"],
"mis_conversions": t["mis_conversions"]
}
for t in dataset["terms"]
if t["priority"] == "最高"
]
# → 53 terms covering life-safety vocabulary
2. RAG Integration
# Build vector DB from safety management terms
safety_terms = [
t for t in dataset["terms"]
if t["category"] == "安全管理"
]
for term in safety_terms:
for context in term["contexts"]:
vector_db.add(
text=context,
metadata={
"term": term["term"],
"priority": term["priority"],
"category": term["category"]
}
)
3. AI Phone Response (Voice Call Automation)
# Map spoken patterns to intent categories
intent_map = {}
for term in dataset["terms"]:
for pattern in term.get("stt_oral_patterns", []):
intent_map[pattern] = {
"term": term["term"],
"category": term["category"],
"priority": term["priority"]
}
# Example: "今日お休みします" ("Taking today off")
# → maps to 欠席連絡 (Absence Notification) category
Data Sample
{
"term": "不適切保育",
"reading": "ふてきせつほいく",
"category": "不適切保育・権利擁護",
"priority": "最高",
"tags": ["行政", "権利侵害", "こども家庭庁", "通報", "保育の質"],
"description": "子どもの人権・尊厳を傷つける保育行為の総称。体罰・暴言・脅し・無視・強制的な排泄・食事の強要など。こども家庭庁が2023年に実態調査を実施し、全国の保育施設に自己点検を義務づけた。",
"mis_conversions": ["不適切補育", "不的切保育", "ふてきせつ保育"],
"stt_oral_patterns": [
"不適切保育があった",
"不適切な対応をしてしまった",
"不適切保育の疑いがある",
"不適切保育として報告する",
"これは不適切保育にあたりますか"
],
"contexts": [
"「不適切保育の疑いがある場合は、一人で抱え込まず管理者に報告してください」",
"「こども家庭庁の指針では、不適切保育を発見した職員には報告義務があります」",
"「行政の実地指導では不適切保育の有無が重点確認事項になっています」"
],
"related_terms": ["体罰", "暴言", "ネグレクト", "権利擁護", "自己点検", "内部通報"]
}
Intended Use Cases
| Use Case | Description |
|---|---|
| STT Custom Dictionary | Improve recognition accuracy for childcare terminology in voice input systems |
| AI Phone Reception | Automated handling of parent calls (absence notifications, pick-up changes) |
| RAG Knowledge Base | Provide domain-specific context retrieval for childcare AI assistants |
| Safety Incident Reporting | Voice-based Hiyari-Hatto (near-miss) reporting systems |
| Regulatory Compliance | Support self-inspection requirements under 2023 CFA guidelines |
| Research | NLP / ASR research in specialized Japanese domains |
Out-of-Scope Use Cases
- General-purpose Japanese NLP (insufficient scale)
- Non-Japanese childcare settings without adaptation
- Medical diagnosis or legal advice applications
Limitations & Future Work
Current Limitations:
- 238 terms is a starting vocabulary, not exhaustive
mis_conversionspatterns are expert-curated, not validated against live STT logs- Oral patterns reflect Tokyo-area childcare settings; regional dialect variation not yet covered
Planned Expansions:
- v2.0: 1,000+ terms with STT log-validated misrecognition patterns
- Regional dialect variations (Kansai, Kyushu, etc.)
- Eldercare (介護) domain extension
- English/Chinese/Vietnamese multilingual support for foreign national families
License
CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)
- ✅ Academic and research use: Free
- ✅ Non-commercial open-source projects: Free
- ✅ Adaptation and redistribution with attribution: Free
- ❌ Commercial product or service integration: Contact us
For commercial licensing inquiries, please contact us at the address below.
Citation
If you use this dataset in research or publications, please cite:
@dataset{infomart_hoiku_stt_2026,
title = {Japanese Childcare Vocabulary Dataset for STT Optimization (保育用語STT最適化データセット)},
author = {Infomart Co., Ltd. (株式会社インフォマート)},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/IDEMITSU/hoiku-yougo-stt-ja},
note = {Built from operational knowledge of Hiyarin, an AI-powered childcare safety reporting system}
}
Contact & Commercial Inquiries
- Product: https://hiyarin.com
- Email: support@hiyarin.com
- Company: Infomart Co., Ltd., Kojimachi Bldg. 4F, 4-2-2 Kojimachi, Chiyoda-ku, Tokyo, Japan
We welcome inquiries regarding:
- Commercial licensing
- Collaborative dataset expansion
- Joint research partnerships
- Integration consulting for childcare AI systems
- Downloads last month
- 60