ketav/hindi-youtube-asr-transcripts

Hindi YouTube ASR Transcripts

Auto-generated YouTube transcripts (VTT) from 21 Hindi channels for training ASR and TTS models.

Quick Start

# Download and extract
wget https://huggingface.co/datasets/ketav/hindi-youtube-asr-transcripts/resolve/main/youtube_asr_data.tar.gz
tar -xzf youtube_asr_data.tar.gz

Stats

Metric	Value
Channels	21
Total videos	109,981
Total hours	22,186.3
Hindi subtitles	98,309
Usable hours	19,032.5
Period	2025-2026
Format	WebVTT with word-level timestamps

Channel Breakdown

Channel	Videos	Total Hrs	Hindi Subs	Usable Hrs
DhruvRathee	722	200.7	565	167.7
JoshTalksHindi	1,695	502.2	1,438	446.4
NDTVIndia	10,000	1,818.0	9,586	1,742.2
News18India	10,000	1,343.7	9,461	1,282.7
NitishRajput	109	48.5	107	48.2
PhysicsWallah	1,377	1,072.9	655	524.3
RanveerAllahbadia	142	201.3	116	163.0
RepublicBharat	10,000	1,757.4	9,762	1,720.5
SansadTV	1,872	509.0	1,405	410.0
SatyaHindi	10,000	1,919.7	9,879	1,892.5
StudyIQEducationLtd	10,000	4,175.1	9,579	3,822.2
ThinkSchool	414	192.7	59	21.4
aajtak	10,000	1,030.8	9,421	987.3
abpnews	1,338	200.8	1,997	192.6
ddnews	10,000	2,575.7	9,564	2,487.2
indiatv	10,000	1,233.6	9,517	1,175.0
khansirofficial	3	0.9	3	0.9
shriprashant	125	109.9	125	109.9
thelallantop	2,184	349.8	2,097	335.0
theprintindia	10,000	1,977.9	3,167	560.2
zeenews	10,000	966.0	9,806	943.3

What's Inside

data/youtube/<channel>/
  videos.json      # Video metadata (id, title, duration)
  manifest.csv     # Maps video_id → transcript files
  subs/
    <video_id>.hi.vtt   # Hindi auto-generated subtitles
    <video_id>.en.vtt   # English (rare)

scripts/
  youtube_download_subs.py   # Download more transcripts
  youtube_scout.py           # Estimate channel hours
  vtt_to_training_data.py    # VTT → audio segments → NeMo manifest

Pipeline: Transcripts → Training Data

# 1. Download audio + segment for one video (dry run)
python scripts/vtt_to_training_data.py --video-id <ID> --channel <handle>

# 2. Download transcripts for more channels
python scripts/youtube_download_subs.py --channel <handle> --workers 6

# 3. List all available channels
python scripts/youtube_download_subs.py --list

VTT Format

YouTube VTT includes word-level timestamps:

00:00:02.560 --> 00:00:05.590
ईरान युद्ध में पूरी दुनिया के लिहाज से
आज<00:00:02.879><c> का</c><00:00:03.120><c> दिन</c>

Quality (from 132 sample analysis)

• Words/sec: 2.75 (normal Hindi: 2-3)
• Hindi chars: 100%
• Auto-caption accuracy: ~85-95% (studio), ~60-75% (field)

Use Cases

• ASR: Train Parakeet-TDT-0.6B or fine-tune Whisper for Hindi
• TTS: Fine-tune IndicF5 or Chatterbox with paired audio
• NLP: Hindi language modeling, code-switching research

Add More Channels

# In scripts/youtube_download_subs.py, add to DEFAULT_CHANNELS:
"NewChannel": {"name": "Channel Name", "category": "news"}

Then run:

python scripts/youtube_download_subs.py --channel NewChannel --workers 6