ytdata/en_50000hrs_clean

English high-quality speech corpus

This release contains the converged cleaned English train manifest, a frozen validation manifest, and a token-coverage-focused subset of approximately 5,000 hours. The existing validated 50,000-hour FLAC/Parquet audio pool is reused without modification or transcoding.

Corpus rows use portable locators of the form audio/<subset>/<group>/part-N.parquet#row=N.

Corpus overview

Corpus/config	Rows	Approx. hours
en/train	10,486,951	49,951.119
en-5000h/train	972,368	4,999.858
validation (shared)	544	1.999

The cleaning audit records exact source and output hashes, rejected audio mappings, language-tag validation, and zero train/validation audio-identity overlap.

Audio subset overview

Counts below are unique packaged audio rows. The audio payload is unchanged by this manifest refresh.

Audio subset	Unique rows	Approx. hours	Rejected rows	Parquet GiB
amphion_emilia_dataset__en	1,281,847	4,460.50	15	388.62
disco_eth_worldspeech__en	307,446	1,622.07	22	134.30
mozilla_common_voice_scripted_english__extracted_audio	3,449	6.91	0	0.66
nvidia_granary__librilight	228,805	1,198.75	0	68.23
nvidia_granary__librilight__extension_50k_v1	1,253,431	5,024.52	0	288.49
nvidia_granary__voxpopuli	662,234	5,236.92	0	324.78
nvidia_granary__voxpopuli__extension_50k_v1	2,307,958	18,218.12	0	1,132.43
nvidia_granary__yodas	2,590,261	10,609.04	776	626.24
nvidia_granary__yodas__extension_50k_v1	1,508,100	2,595.48	0	156.32
nvidia_granary__ytc	2,426	9.26	0	0.55
voices_in_the_wild_jsonl__audio	355,155	1,016.17	0	73.52

JSONL examples

{"text":"Yes, we're constantly giving away lots of free merchandise. Automotive Diagnostics and Publishing's, Mandy Concepcion.","duration":7.658,"context":"We present all the content using the latest CG animation techniques on hand's video, and how to, tips and techniques. We encourage you to subscribe to this channel now. Once subscribed, any time we upload a new automotive tip, secret, or technology video, you will be notified. Finally, by subscribing, you will also be part of our weekly freebies.","lang":"<|EN|>","audio_id":"9e4baf0515be3261eea308e74f698dea78af6602","source_set":"amphion_emilia_dataset__en","audio_filepath":"audio/amphion_emilia_dataset__en/2e730258c31aa0c8311b/part-00000.parquet#row=998","audio_parquet":"audio/amphion_emilia_dataset__en/2e730258c31aa0c8311b/part-00000.parquet","audio_row":998,"audio_format":"flac"}
{"text":"This week in tech is brought to you by Rocket Mortgage from Quicken Loans. When it comes to the big decision of choosing a mortgage lender, work with one that has your best interest in mind. Use Rocket Mortgage for a transparent, trustworthy home loan process that's completely online at QuickenLoans.com slash twit2.","duration":17.372,"context":"This is twit, this week in tech, episode 599, recorded Sunday, January 29th, 2017. Fold with dignity.","lang":"<|EN|>","audio_id":"424bfed03c59e8c7d6b5df8c5c1dcade4d7a7c55","source_set":"amphion_emilia_dataset__en","audio_filepath":"audio/amphion_emilia_dataset__en/51690e5e849db935410c/part-00000.parquet#row=226","audio_parquet":"audio/amphion_emilia_dataset__en/51690e5e849db935410c/part-00000.parquet","audio_row":226,"audio_format":"flac"}

Loading

from datasets import load_dataset
full = load_dataset("ytdata/en_50000hrs_clean", "en", split="train")
coverage = load_dataset("ytdata/en_50000hrs_clean", "en-5000h", split="train")
validation = load_dataset("ytdata/en_50000hrs_clean", "en", split="validation")

To download only selected source audio, use snapshot_download with allow_patterns for corpus/* and the required audio/<subset>/** paths.

Audio release totals

• Unique audio: 10,501,112
• Audio hours: 49,997.746
• Parquet size: 3.429677 TB (3.119273 TiB)

See release.json for exact per-subset counts and corpus checksums, and cleaning_audit.json for refresh validation details.