CashlessConsumer/tamil-cc-yt-captions

Tamil CC YouTube Captions (தமிழ் CC வலையொளி வரைவுகள்)

YouTube videos marked Creative Commons by YouTube (license filter videoLicense=creativeCommon in the official YouTube Data API v3 search.list), with their Tamil captions. Built for Tamil speech-recognition training (e.g. fine-tuning udgamlabs/Shruti) and for Tamil NLP corpus work.

Harvest date: 2026-11-05 · Collector: CashlessConsumer (github.com/CashlessConsumer)

What is in here

	count
CC-licensed Tamil videos listed (data/listing.jsonl)	493
With Tamil captions fetched (data/metadata.jsonl + captions/)	100
Listed audio duration	57.1 h
Captioned audio duration	32.6 h
Channels	99

• data/listing.jsonl — every CC video found: id, url, title, description, channel, publish date, duration, has_ta_captions, ta_caption_kind (auto/manual/null).
• data/metadata.jsonl — the subset with downloaded captions; adds caption_file.
• captions/<videoId>.ta.vtt — raw YouTube caption files, timestamped.

License & provenance (read before using)

1. Video license is CC-BY 3.0 as declared by YouTube. Every row was returned by YouTube Data API v3 search.list with videoLicense=creativeCommon — the license mark is YouTube's, queried on 2026-11-05. CC-BY requires attribution: credit the channel, link the video URL, and state edits (this dataset card's metadata does this per row; downstream users must propagate it).
2. Captions are NOT producer-uploaded. 45 of the 46 Tamil caption tracks are auto-generated by YouTube's ASR (field ta_caption_kind: "auto"); exactly one video (vSqfXNvqSlg, Pesum Tamil) has producer-written captions ("manual"). Treat all caption text as silver: expect word-boundary errors, missing punctuation conventions, and ASR mishears — clean before claiming benchmark numbers. The caption text itself carries no separate YouTube license mark; it is distributed here as derived metadata with full provenance, alongside the CC-BY video it describes.
3. Audio is not included. To pair audio with captions, download each video with yt-dlp from a non-datacenter IP (Google Colab works) and resample to 16 kHz mono:

# Colab-ready: fetch audio for this dataset's videos
import json, subprocess, os
!pip -q install yt-dlp soundfile
!apt -qq install -y ffmpeg

import urllib.request
URL = "https://huggingface.co/datasets/CashlessConsumer/tamil-cc-yt-captions/resolve/main/data/metadata.jsonl"
rows = [json.loads(l) for l in urllib.request.urlopen(URL)]
os.makedirs("wav", exist_ok=True)
for r in rows[:20]:  # start small; ~32 h total
    vid = r["videoId"]
    if not os.path.exists(f"wav/{vid}.wav"):
        subprocess.run(["yt-dlp", "-x", "--audio-format", "wav",
                        "--postprocessor-args", "ffmpeg:-ac 1 -ar 16000",
                        "-o", f"wav/{vid}.%(ext)s", r["url"]], check=False)
print("done", len(os.listdir("wav")))

4. Channel mix skews to explainer/news/interview speech (Karka Kasadara, PuthiyathalaimuraiTV, Tamil Speech and Story, DMK IT Wing, IAS academies, …) — formal, read-style Tamil dominates; spontaneous conversational speech is underrepresented. Don't use as a general-domain benchmark.

Method

• Query set: 6 topical queries (speech, audiobook, education, interview, story, lecture) + per-channel scans (channelId + videoLicense=creativeCommon, order=date) across the 99 channels discovered in round 1. YouTube Data API v3.
• Captions listed via Invidious /api/v1/captions/<id>, filtered to ta, fetched as WebVTT, kept verbatim.
• Collector scripts: ThamizhKanimai/nlp/shruti-tamil/yt-data (harvest_captions.py, expand_channels.py).

Known limits

• The 393 listed-but-not-yet-captioned videos: caption fetching is rate-limited on public Invidious instances; ids are ready in data/listing.jsonl.
• durationSec comes from Invidious metadata; a few may be stale (live streams excluded via liveNow filter at harvest).
• Copyright: the CC mark is YouTube's per-video declaration; if a rights holder disputes a video's CC status, that row should be dropped — raise an issue.