Vietnamese YouTube Speech Corpus — yt2026_batch02_10kh

🇻🇳 Tiếng Việt nói thật, thu từ YouTube — chỉ video xuất bản từ năm 2026 trở đi. Dữ liệu có tiếng, không kịch bản, để huấn luyện và đánh giá ASR tiếng Việt trong điều kiện thực tế.

🇬🇧 Real, un-scripted Vietnamese speech harvested from YouTube — 2026 uploads only. Raw audio for training and evaluating Vietnamese ASR under real-world conditions.

Generated 2026-10-06 · one unit = ~10,000 h · more units on the way toward 50k–100k h.

🗓 2026-only by construction — this unit does not overlap older corpora

Every clip here was published on or after 2026-01-01. That is a hard filter applied at crawl time, not a sample that happened to look recent, and the field is auditable per clip: upload_date in manifest.jsonl.

If you are looking for Vietnamese speech, this is the 2026-onward slice — it is disjoint from public corpora covering earlier material (GigaSpeech 1 covers 2018–2021, GigaSpeech 2 spans pre-2026 releases). Nothing in this repository is duplicated from those releases.

At a glance

Hours of audio	11,884.0 h
Audio files	17,754 (.webm, Opus 48 kHz mono, no re-encoding)
Channels	320 YouTube channels
Metadata files	18,800 info.json (yt-dlp; 36,554 files total)
Total size	624.7 GB
Publication window	20260101 → 20261006 — 2026-only (hard filter ≥ 20260101)
Median clip	613 s (min 2 s · max 81109 s)
Clips ≤ 10 s / ≥ 30 min	166 / 5,853
Mean / median file	35.2 MB / 9.1 MB

Why this corpus

Public Vietnamese speech corpora are small and skew towards clean read speech. Real users listen to commentary, news, religious talks, street interviews and live streams — with accents, code-switching to English, background noise and microphone artefacts.

This corpus is built for that gap:

• Un-scripted, conversational speech instead of read sentences.
• Vietnamese ↔ English code-switching happens naturally in modern Vietnamese media.
• Full acoustic range: phone calls, radio/TV, field recordings, screen-capture audio.
• Documented provenance per clip (video_id, channel_id, upload_date) so any item can be audited, reproduced or removed.
• Disjoint from older corpora. Restricting the crawl to 2026+ uploads means this slice does not duplicate the 2020–2025 Vietnamese material already covered elsewhere — useful when you combine sources without worrying about overlap.

What's inside

audio/<channel_id>/<upload_date>#<title>#<channel_id>#<video_id>_<duration>.webm
audio/<channel_id>/<upload_date>#<title>#<channel_id>#<video_id>_<duration>.info.json
manifest.jsonl   ·  one JSON object per audio file
SHA256SUMS       ·  sha256sum -c compatible
files.txt        ·  plain list of all relative paths
stats.json       ·  totals, per-channel and per-month aggregates
used_ids.txt     ·  video_ids in this unit (frozen: no duplicates across units)
UNIT.json        ·  unit metadata and layout contract

manifest.jsonl fields

field	meaning
rel_path	path relative to audio/
audio_path	path relative to the crawl root
video_id / channel_id	YouTube IDs (join with https://www.youtube.com/watch?v=<video_id>)
upload_date	YYYYMMDD — the enforced 2026 window
duration	seconds
size / mtime	bytes / unix mtime
sha256	integrity hash
info_json_path	matching yt-dlp metadata file

How to use it

Stream/download one file

from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="SalmonAI123/yt2026_batch02_10kh", repo_type="dataset",
                     filename="UC.../20260930#...#<video_id>_<dur>.webm")

Bulk download (filtered)

hf download SalmonAI123/yt2026_batch02_10kh --repo-type dataset --include "UCxxx/*" --local-dir ./subset

Verify integrity

sha256sum -c SHA256SUMS        # every file, sha256

Load the metadata

import json
rows = [json.loads(l) for l in open('manifest.jsonl')]
print(len(rows), 'clips,', sum(r['duration'] for r in rows)/3600, 'hours')

This repo is a raw-audio release (not a datasets-loadable build). Fetch files with hf_hub_download / hf download, then build your own shards (tar/webdataset/parquet) locally.

Coverage

month	clips	hours
202601	696	354.1
202602	471	248.1
202603	776	316.2
202604	914	384.7
202605	1,245	585.4
202606	1,304	754.2
202607	1,603	873.3
202608	2,811	1,766.1
202609	5,721	5,187.6
202610	2,213	1,414.3

Quality & limitations

• No transcripts. This is the acquisition stage: audio + provenance only.
• 2026-only, by design. Nothing published before 2026-01-01 is included. Verify per clip with upload_date in manifest.jsonl.
• Selection bias. Channels were harvested from a Vietnamese-language discovery list; it is not a demographically balanced sample.
• Channel-level licensing varies. Items inherit whatever rights their original creator holds; see Terms of access.
• Content may change. A channel owner can delete a video, in which case that item may be removed here after a notice.
• Partially downloaded (*.part) files are excluded by construction.

Ethics, provenance & Terms of access

This corpus is acquired by automated crawling of public YouTube for non-commercial research on Vietnamese speech. By downloading or using it you agree to:

1. Use it for research only — not for commercial products or services.
2. Do not redistribute it (or derived audio) as your own dataset or model release.
3. Comply with YouTube's Terms of Service and with the copyright law that applies to you.
4. Cite this corpus (see below) and this repository.

Provenance. Every clip carries its video_id, channel_id, upload_date and the crawl timestamp (2026-10-06); the original URL is reconstructible as https://www.youtube.com/watch?v=<video_id>.

Takedown. Rights holders may ask for removal of specific items, whole channels, or the entire dataset. Requests are honoured — contact the repository owner and we will act within 7 days. Because everything is keyed by video_id, removals are precise and reproducible.

Citation

@misc{yt2026_batch02_10kh,
  title  = {Vietnamese YouTube Speech Corpus (yt2026) — raw audio for Vietnamese ASR},
  note   = {Unit 10kh, 11,884 hours, 17,754 clips, 320 channels},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SalmonAI123/yt2026_batch02_10kh}

Roadmap

• Transcribe (Whisper-family models) + forced alignment, released as separate repos.
• More units toward 50k–100k hours; each unit is versioned, checksummed and gated.

---

Generated by metadata/make_hf_card.py — do not edit by hand; re-run the generator instead.