Tabahi/yt-cc-speakers

YouTube CC Speaker Corpus

English speech from YouTube, organized by speaker: 13,051 recordings and 3,716 hours from 49 speakers. Each recording is the audio of one complete YouTube video, denoised so that only the voice remains, with an automatically generated transcript and the video's YouTube metadata. Each speaker is the main voice of one YouTube channel.

This is the full corpus. A more balanced six-speaker dataset with a train/test split is published separately and more openly as yt-cc-speakers-6spk. The related experiment paper is at: arxiv.org/abs/2610.05575

All recordings come from YouTube videos that were published under the Creative Commons license (CC BY 3.0). See Source and license.

At a glance

Recordings	13,051
Speakers (channels)	49
Total audio	3,715.9 hours
Female speakers	22 speakers, 4,677 recordings, 1,634.0 hours
Male speakers	27 speakers, 8,374 recordings, 2,081.9 hours
Recording length	11 s to 4.3 h, median 13.2 min
Videos published	2011 to 2025
Language	English
Audio format	FLAC, 16 kHz, mono, 16-bit

There is no train/test split: this is the complete set of recordings. The speaker IDs are not consecutive (some numbers between 0 and 53 are unused).

Speakers and voices

The channels were chosen because one person does most of the talking, and a speaker ID stands for one channel and its main voice. This is not a conversational corpus: most recordings are that one person speaking alone.

Some recordings do contain other voices, such as a guest, a co-host, or a discussion between several people. The data has no per-segment speaker labels, so those voices are not marked, and everything in a recording is filed under the channel's speaker ID.

Known limitations

• Transcripts are automatic. They were not checked by a person and contain recognition errors, including in names.
• Other voices are not labeled. See Speakers and voices.
• Denoising artifacts. Removing noise and music can leave gaps or distortion where the voice overlapped with other sound.
• No recordings without speech. Recordings whose transcript came out empty (short clips in which no speech was recognized) are excluded.

Loading

from datasets import load_dataset

ds = load_dataset("Tabahi/yt-cc-speakers", split="train", streaming=True)
row = next(iter(ds))
row["audio"]        # decoded audio, 16 kHz
row["text"]         # full transcript
row["speaker_id"]

The dataset has no train/test division; train is only the name the Hub gives to a single set. Each row is one whole recording, and recordings can be hours long, so streaming is the practical way to iterate.

metadata.jsonl holds one row per recording:

Column	Content
file_name	Path to the audio file (loaded as the audio column)
speaker_id, speaker_name, gender	Speaker ID, channel name, and gender label
video_id, channel_id	YouTube identifiers
title, published_at, source_url	Title, publication time, and link of the source video
duration	Length of the audio in seconds
text	Full transcript
transcript_file	Path to the transcript JSON with segment and word timestamps
metadata_file	Path to the video's metadata JSON

Layout

yt-cc-speakers/
    README.md               this file
    metadata.jsonl          one row per recording, used by load_dataset
    ATTRIBUTION.csv         credit line for every recording
    speakers.csv            one row per speaker
    all.json / all.csv      index of every recording
    missing_sources.csv     selected recordings that are not included, with the reason
    organization.json       record counts
    speaker_<id>/
        speaker_<id>.json / .csv    index of this speaker's recordings
        audios/<videoId>.flac
        captions/<videoId>.json
        meta/<videoId>.json

speakers.csv columns:

Column	Content
speaker_id	Speaker ID, used in the directory names and in spkr
name, channel_title	Name of the speaker's YouTube channel
channel_id	YouTube channel ID
gender	Gender label, M or F
lang	Locale label recorded for the channel during collection
videos, total_hours	Recordings in this dataset and their total length, measured from the audio
videos_listed, hours_listed	Recordings originally selected for the speaker
videos_missing	Selected recordings that are not included

631 of the originally selected recordings are not included: 470 because their audio or transcript was unavailable, 152 because no speech was recognized in them (short clips with an empty transcript), and 9 because YouTube flags the video as made for kids. They are listed in missing_sources.csv with the reason.

Speakers

ID	Name	Channel ID	Gender	Locale	Recordings	Hours
0	The Aussie Flipper	UC-GfooZBv2_IEGxhv4SoleQ	M	en-AU	545	164.28
2	Kenzie Elizabeth	UC-mConRSkyYyFrvMH3zQNMQ	F	en-US	524	197.18
3	Cincinnati Picker	UC1e7o39iNBeLke6tDaykN2A	M	en-US	457	108.76
5	Melvin and Dane	UC3oQ_i0ovVuntRLwfmDMwsA	M	en-US	27	10.38
6	Freya Farrington	UC3w9E1N7zukLxJ5bkrU4IKw	F	en-GB	430	185.90
7	Dr. Andrew Neville	UC6XlVIx6XfBrQaNWasTNllQ	M	en-US	235	24.25
8	Alex The Analyst	UC7cs8q-gJRlGwj4A8OmCmXg	M	en-US	318	129.87
9	Nazje Noir	UC978AgSRPAAzewp5oiR_S4Q	F	en-US	106	52.61
10	Jenn Clayton	UCALeORBToC1O0BtmE7Xr1bg	F	en-US	666	199.95
11	Daily Carrington	UCGFR38av-AtLe7xGsKA9KDw	F	en-GB	18	7.33
12	vlogbrothers	UCGaVdbSav8xWuFWTadK6loA	M	en-US	1,010	81.43
13	The Dream Green Show	UCGvHysu4wXjPRXIjEn3NEGw	M	en-US	474	90.82
15	The Mindset Mentor Podcast	UCHl3aFKS0bY0d8JwqNysaeA	M	en-US	378	117.34
16	Drinkin' Bros Studios	UCLK5icWTfKgw6I2HIku84vg	M	en-US	186	200.04
17	Cleetus2 McFarland	UCMJbbWyIziiUamP5JKK9yAA	M	en-US	148	40.66
18	Practical Engineering	UCMOqf8ab-42UUQIdVoKwjlQ	M	en-US	6	0.71
19	Carrie Dayton Vlogs	UCMXBBb6GLVJI8-Lie7QAMkQ	F	en-US	144	45.72
20	Cinematic Excrement	UCStzEOk2Pfftjy2f2sHjflg	M	en-US	484	81.62
22	christine lee	UCUzy4w3TfWgoXTLXyXwEbHw	F	en-US	64	17.01
23	Terry and The Guide Inside	UCXcADyv8rVgc925auBgaENw	M	en-??	158	11.23
24	Morgan Bylund	UCYBGUkYwMsxkc-P7Z8Cw7IA	F	en-US	327	104.15
25	Getting It Done - Budget Home Repairs	UC_C2O89kMQuDStUEoM6Fz1w	M	en-US	108	12.58
27	Shifting Metal	UCaz71xNEW8qdTMzhG5_rMGg	M	en-GB	287	85.86
28	The Budget Minded Motorist	UCb38N20J0JMqfwMgRMX-fBg	M	en-US	4	0.37
29	Akeira Janee	UCbAIYQeu0qxmfYJHpo1ZAkA	F	en-US	193	86.96
30	Emily Lucy Rajch	UCbdITXvPvEmeTnwSPrUnn3A	F	en-GB	107	24.70
31	Scott Ste Marie	UCciA8PsPm3Z5E5vlKEct_lg	M	en-CA	153	29.83
32	Anton Petrov	UCciQ8wFcVoIIMi-lfu8-cjQ	M	en-CA	2	0.55
33	Jasmyn Destine	UCdwQH9oe8c5t48jE5X5KNYA	F	en-US	49	20.34
34	Crossthread Garage & Salvage	UCed-0rjG9PEGK-xpTynM8oA	M	en-US	21	14.05
35	Mike Day Concrete	UCem87zDMe98Cmypjd6PVf9A	M	en-US	1,444	187.64
36	Rose Keats	UCeo16X9mvT_8UCb1QsWhHHQ	F	en-IR	78	21.02
37	Faith Elizabeth Vlogs	UCf9tMCeiq0k0koGvBAIZF2w	F	en-US	122	20.08
38	ashleybchoi	UCfB2E_culXfbKpxr7KqzV3Q	F	en-US	129	38.20
39	ZachCowman	UCfcyMOhYjJGp1Tie3kqvFyw	M	en-US	44	27.89
40	Paige Lorenze	UCiBm-VyAD9rud1xBgyl-fUQ	F	en-US	43	15.12
41	Danielle Marie Carolan	UCiGWpX9oCmMu3hRUV_ZBvpg	F	en-US	557	179.93
42	Michaela Bento	UCjtmwPW5sB95RSnFJzPKzAQ	F	en-CA	234	77.46
43	Anything Goes With James English	UCkdiBNdMSiQeT8aD7gXWgvA	M	en-GB	265	198.23
44	Adam Does Movies	UCmuP70--NYoqgyo3N0ZwDCA	M	en-US	765	199.67
45	Kirah Ominique	UCpUtNtqyGxuh3cmfSkMrJMQ	F	en-US	311	165.34
46	Jordan Bauth	UCqkbECaAqs-mpOC09-R4eRw	F	en-US	153	42.14
47	Trucker Josh VLOGS	UCrkahiSmFd6w0fmdZ95K_wA	M	en-CA	601	199.69
48	Alyssa Jalise	UCuRYtgM907DhjrIvgzhENAg	F	en-US	148	49.18
49	Petrolhead Rebuilds	UCwi2UtTEegtzB4OT1WZDChw	M	en-GB	30	9.32
50	We Might Be Drunk Podcast	UCy6A9WMN43DrtBkID7nMXJw	M	en-US	13	19.49
51	Persian Bunny	UCyiCwUmJ_l4Vrzur70OqTpw	F	en-GB	210	64.20
52	Neil Mossey	UCz21nEmP8HLh6VqqQdlCvJQ	M	en-GB	211	35.39
53	Diana Malloy	UCzYy8Ax-rFSXOEgqRmy3AmQ	F	en-US	64	19.45

File formats

Audio (audios/<videoId>.flac): FLAC, 16 kHz, mono, 16-bit. One file per source video, the full length of the video. The audio has been denoised with the Facebook Research Denoiser (https://github.com/facebookresearch/denoiser): only the voice is kept, and background noise and music are removed.

Transcripts (captions/<videoId>.json): machine-generated transcripts from Whisper large-v3-turbo, with segment and word timestamps. The files are JSON, although the index field that points to them is called srt for historical reasons.

Key	Content
text	Full transcript of the recording
segments	List of segments, each with start, end (seconds), text, and a words list with per-word start, end, and probability. Segments also carry decoder fields such as tokens, avg_logprob, and no_speech_prob
language	Detected language (en)
ori_dict	The transcription result before segments were regrouped
nonspeech_sections	start/end times of detected non-speech (not present in every file)
regroup_history	Regrouping settings used (not present in every file)
audio_path	Path of the recording's audio file, relative to the directory of the index file (speaker_<id>/audios/<videoId>.flac)

Metadata (meta/<videoId>.json): YouTube metadata for the source video.

Key	Content
videoId, channelId, channelTitle	YouTube identifiers and the channel name
title, name	Video title (name repeats it)
description	Video description as written by the creator, with links and email addresses removed
publishedAt	Publication time (UTC)
duration	Duration in seconds as reported by YouTube
definition	hd or sd for the source video
license	Always creativeCommon
madeForKids	YouTube's audience flag; always false, because made-for-kids videos are excluded
viewCount, likeCount, commentCount	Statistics at collection time (some files only)
downloaded, chunks	Collection bookkeeping (some files only)

Index records (*.json and *.csv index files): one entry per recording.

Field	Content
audio	Path to the audio file, relative to the directory of the index file
srt	Path to the transcript file, relative to the directory of the index file
name	Video title
duration	Duration in seconds as reported by YouTube
audio_duration	Duration in seconds measured from the audio file
gender	Speaker gender label, M or F
spkr	Speaker ID
channel_id	YouTube channel ID
videoId	YouTube video ID

Source and license

Every recording in this dataset comes from YouTube, and only from YouTube. A video was included only if YouTube reported it as published under the Creative Commons license at collection time. That check is recorded for every video: each meta/<videoId>.json file carries "license": "creativeCommon", and there are no records in this dataset with any other value.

YouTube offers creators exactly one Creative Commons option, the Creative Commons Attribution 3.0 Unported (CC BY 3.0) license:

• Summary: https://creativecommons.org/licenses/by/3.0/
• Legal code: https://creativecommons.org/licenses/by/3.0/legalcode

Under CC BY 3.0 you may copy, redistribute, remix, transform, and build upon the material for any purpose, including commercially, as long as you give appropriate credit. The license does not expire and cannot be revoked for material obtained while it applied, but a creator may stop offering a video under it, or remove the video from YouTube, at any time.

Attribution

The copyright in each recording belongs to the creator of the source video, not to the dataset maintainers. ATTRIBUTION.csv gives the credit line for every video, one row per recording:

Column	Content
spkr	Speaker ID used in this dataset
videoId	YouTube video ID
title	Title of the source video
creator	YouTube channel that published the video (the author to credit)
source_url	Link to the source video
channel_url	Link to the creator's channel
published_at	Publication date of the source video
license, license_url	CC BY 3.0 and the link to the license
audio, captions	The files in this dataset derived from that video

If you redistribute any of this data, or publish work that reproduces it, you must:

1. Credit the creator and give the title of the video.
2. Link to the source video.
3. Name the license (CC BY 3.0) and link to it.
4. State that changes were made (see below).
5. Not suggest that the creator endorses you or your use.
6. Not add legal terms or technical measures that stop others from doing what the license allows.

A credit in this form satisfies points 1 to 4:

"We Started An eBay Business From Scratch" by The Aussie Flipper (https://www.youtube.com/watch?v=SYfbV6_SjOU), licensed under CC BY 3.0 (https://creativecommons.org/licenses/by/3.0/). Audio extracted, converted to 16 kHz mono, and denoised; transcript generated automatically.

When you use many recordings at once, citing this dataset together with its ATTRIBUTION.csv is a reasonable way to give that credit.

Changes made to the source material

• Only the audio track of each video is included; there is no video.
• The audio was converted to 16 kHz, mono, 16-bit.
• The audio was denoised with the Facebook Research Denoiser (https://github.com/facebookresearch/denoiser) so that only the voice remains. Background noise and music, including music the creator added, have been removed. The recordings therefore do not sound like the source videos.
• The denoised audio is stored as FLAC, which is lossless.
• The transcripts were produced by automatic speech recognition with Whisper large-v3-turbo. They are not the creators' own captions and contain recognition errors.
• Video titles, descriptions, and statistics are reproduced as YouTube metadata in meta/. Links and email addresses have been removed from the descriptions.

Other things to know

• The license covers copyright only. It does not cover the privacy or publicity rights of the people who speak or are mentioned in a recording, nor trademarks, nor third-party material (music, clips) that a creator may have included in a video.
• The license check reflects the state of each video when it was collected.
• Videos that YouTube flags as made for kids are excluded.

Citation

@misc{yt_cc_speakers_2026,
    author       = { Abdul Rehman },
    title        = { yt-cc-speakers (Revision 378d8c8) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/Tabahi/yt-cc-speakers },
    doi          = { 10.57967/hf/10891 },
    publisher    = { Hugging Face }
}

Contact and removal requests

The dataset maintainer can be reached at e420@duck.com. If you are the creator of a source video and want a recording removed or a credit corrected, write to that address with the video link.

• The material is provided as is, without warranties.