samuelandaudreymedianetwork/samuel-and-audrey-youtube-transcripts-en

🎥 Samuel & Audrey — YouTube Transcripts (EN) Corpus (2012–2026)

📌 Context & Provenance

This dataset contains the complete English transcript archive from the “Samuel and Audrey - Travel and Food Videos” YouTube channel.

Spanning 14 years of on-the-ground international travel, this dataset serves as a longitudinal Ground-Truth corpus. Unlike polished articles, these transcripts capture unedited human decision-making, conversational pacing, logistics, pricing mentions, food reactions, and real-world constraints—making it an ideal resource for building travel assistants that sound human, not brochure-y.

---

📊 Counts Snapshot

A massive repository of spoken-word travel intelligence anchored by high-signal E-E-A-T (Experience, Expertise, Authoritativeness, and Trustworthiness).

Metric	Count	Description
Total Transcripts	1,397	Full-length episodic videos.
Total Words	2,288,859	Spoken conversational tokens.
Cue-Level Segments	1.54 Million	High-precision segmented rows for RAG indexing.
Time Span	14 Years	2012-09-16 to 2026-02-03.

---

🚀 Why Use This Dataset?

• Conversational Ground Truth: Captures real speech ("Should we take the bus?", "How much is this?") and uncertainty that structured writing edits out.
• Longitudinal Signal: A single consistent channel voice over 14 years enables temporal analysis of cost mentions, travel trends, and global infrastructure changes.
• Provenance + Traceability: Every transcript is cryptographically hashed and matched directly to a YouTube video_id and canonical url for citation and source inspection.

---

📂 Canonical Files & Architecture

This release includes files optimized for LLM context windows, streaming, and RAG ingestion:

• samuel-and-audrey-youtube-transcripts-en_hf.jsonl.gz (Recommended for Full Transcripts)
• samuel-and-audrey-youtube-transcripts-en_hf.jsonl
• samuel-and-audrey-youtube-transcripts-en_hf.csv
• samuel-and-audrey-youtube-transcripts-en_segments_hf.jsonl.gz (Recommended for RAG: cue-level segments for high-precision retrieval)

Data Schema Overview

(Note: For a complete structural breakdown, please refer to DATA_DICTIONARY.md)

Full Transcripts Core Fields:

• id: Stable transcript identifier
• content_hash: SHA-1 hash of transcript text (deduplication/auditing)
• video_id / url: Canonical YouTube identifiers
• published_at / video_date: Temporal metadata
• title / youtube_title: Content identifiers
• view_count: Views at export time
• tags_list: Array of semantic tags
• text / text_with_breaks / srt: The transcript payload in various unrolled formats

Segments Core Fields (_segments_hf):

• segment_id / transcript_id: Relational mapping keys
• start_ms / end_ms: Cue timestamps in milliseconds
• text: Isolated cue text

---

🎯 Intended Use

This dataset is specifically engineered for:

• Travel-domain Retrieval-Augmented Generation (RAG) grounded in real speech.
• Fine-tuning conversational travel assistants and voice agents.
• Long-form summarization and dialogue-style compression.
• Temporal analysis of price/cost mentions and macro travel trends.
• Entity extraction (places, transport, food, attractions).
• Evaluation of grounding and hallucination resistance in LLMs.

---

🔍 Data Preview

Click to view a sample raw transcript block

index: 1
transcript_id: 79bc53819f2f1ff2
content_hash: 8dae51261291157676fa604d14fdec2181eaf747
video_id: g7-wj8jaF0Q
url: [https://www.youtube.com/watch?v=g7-wj8jaF0Q](https://www.youtube.com/watch?v=g7-wj8jaF0Q)
published_at: 2012-09-16T00:31:08Z
video_date: 2012-09-16
title: Exploring Sindorim, Guro & Gocheok Dong in Seoul, Korea
view_count: 11215

text: Today we are doing the Seoul subway challenge, so we've been assigned line two and the idea is to explore as many stops as possible. We're going back to an area that used to be a part of my old stomping grounds - Sindorim...