SozAI keeps the YouTube transcripts that visitors request on sozai.app as public pages, and on 24 October 2026 that library held 27,852 published transcripts. Nobody planned them as a sample. Almost every one exists because someone pasted a link and wanted the text, so the library is a record of which videos people wanted to read. This article goes through what is in it: languages, length, topics, channels and growth since February.
Data snapshot: 24 October 2026, 05:13 UTC. Source: the public transcript library on sozai.app.
The library at a glance
| Measure | Value |
|---|---|
| Published transcripts | 27,852 |
| Distinct YouTube videos | 27,838 |
| Languages | 84 |
| Distinct channels | 15,515 |
| Total video time | about 15,189 hours |
| Median transcript length | 2,121 words |
| Median video length | 16.3 minutes |
Visitors on the website asked for 25,668 of the transcripts. The SozAI browser extension accounted for 383, and the study tools on the site for 142. We added 680 ourselves as a medical-education collection. Another 979 older posts have no source recorded. A video gets a transcript once, and the next person who asks for it is sent to the existing page. That is why the number of distinct videos is almost the same as the number of transcripts: only 14 videos have two pages.
Languages
A transcript’s language is the language of the caption track it was built from. Here are the 15 largest:
| Language | Transcripts | Share |
|---|---|---|
| English | 10,247 | 36.8 % |
| Russian | 6,024 | 21.6 % |
| Spanish | 1,429 | 5.1 % |
| Arabic | 1,281 | 4.6 % |
| Polish | 1,136 | 4.1 % |
| Hindi | 1,037 | 3.7 % |
| Indonesian | 896 | 3.2 % |
| French | 745 | 2.7 % |
| Ukrainian | 630 | 2.3 % |
| Portuguese | 591 | 2.1 % |
| Korean | 501 | 1.8 % |
| Japanese | 482 | 1.7 % |
| German | 416 | 1.5 % |
| Chinese | 329 | 1.2 % |
| Italian | 315 | 1.1 % |
| 69 other languages | 1,793 | 6.4 % |
English accounts for just over a third of the library, which leaves almost two thirds in other languages. The 15 largest languages cover 93.6 % of all transcripts. The other 69 share the remaining 6.4 %. Russian is a strong second at more than one transcript in five. Below the top 15 the long tail runs from Tamil (221) and Turkish (197) down to languages with a single transcript.
The English figure is slightly too high. Before 20 October 2026, some videos in other languages were saved with their English auto-dubbed captions in place of the original track. About 930 published posts have a title in a non-Latin script but are labelled English. A separate article covers how that happened and what changed.
Length
The median transcript has 2,121 words, about what you would read in a long magazine feature. The spread is wide:
| Words per transcript | Transcripts | Share |
|---|---|---|
| Under 500 | 6,136 | 22.0 % |
| 500 to 1,999 | 7,333 | 26.3 % |
| 2,000 to 4,999 | 7,511 | 27.0 % |
| 5,000 to 9,999 | 4,057 | 14.6 % |
| 10,000 or more | 2,815 | 10.1 % |
A quarter of the transcripts are under 629 words. A quarter are over 4,931, and the middle half lies between those two figures. One in ten passes 10,076 words. The longest has 627,001.
Video length follows the same pattern:
| Video length | Transcripts | Share |
|---|---|---|
| Under 5 minutes | 5,701 | 20.5 % |
| 5 to 14 minutes | 7,445 | 26.7 % |
| 15 to 29 minutes | 6,298 | 22.6 % |
| 30 to 59 minutes | 4,563 | 16.4 % |
| 1 to 2 hours | 2,640 | 9.5 % |
| 2 hours or more | 1,204 | 4.3 % |
The median video runs 16.3 minutes, and the middle half falls between 6.7 and 36.9 minutes. 13.8 % of the videos (3,837) are longer than an hour, and 30.2 % are longer than half an hour. The mean transcript (4,307 words) is about twice the median, because those long videos pull it up.
Topics
On 30 July 2026 we sorted every transcript published up to that date into 20 topics. 13,008 of the published transcripts carry a topic. Later transcripts have not been sorted yet, so the shares below are of the tagged set.
| Topic | Transcripts | Share of tagged |
|---|---|---|
| Science & Education | 2,130 | 16.4 % |
| Psychology & Self-Development | 1,048 | 8.1 % |
| Business & Startups | 894 | 6.9 % |
| News & Politics | 870 | 6.7 % |
| Music | 842 | 6.5 % |
| Tech & AI | 841 | 6.5 % |
| Religion & Spirituality | 763 | 5.9 % |
| Film & TV | 631 | 4.9 % |
| Software How-To | 631 | 4.9 % |
| Health & Fitness | 596 | 4.6 % |
| Money & Investing | 584 | 4.5 % |
| Lifestyle | 461 | 3.5 % |
| Society & Culture | 459 | 3.5 % |
| Gaming | 449 | 3.5 % |
| Language Learning | 438 | 3.4 % |
| Stories & Audiobooks | 388 | 3.0 % |
| Astrology & Esoterica | 355 | 2.7 % |
| Career & Job Search | 328 | 2.5 % |
| Sports | 178 | 1.4 % |
| Food & Cooking | 122 | 0.9 % |
Science and education is the largest topic by a wide margin, about twice the size of the next one. Eight topics are about learning or work: science and education, psychology, business, tech, software how-to, money, language learning and careers. Together they hold 53.0 % of the tagged transcripts. Sports (1.4 %) and food (0.9 %) are the smallest topics.
Channels
The 27,852 transcripts come from 15,515 channels. Most channels appear only once: 79.6 % of them have exactly one transcript in the library, and 264 have ten or more. No channel dominates. The ten largest channels requested by visitors together make up about 3 % of the library:
| Channel | Transcripts |
|---|---|
| Артём Звёздин – обучение трейдингу | 145 |
| Umiejętności Jutra AI | 119 |
| Shivanshu Agrawal | 89 |
| Retro Aanmeegam | 72 |
| Paweł Paczul | 72 |
| Mundo Tesis (Oficial) | 70 |
| Réviser la Biochimie With Me | 69 |
| Chris ‘The Wiz’ Miller | 68 |
| TEDx Talks | 66 |
| Alumni Fund Ukraine | 66 |
This list leaves out four medical-education channels from the collection we added ourselves. Each of them has between 74 and 174 transcripts.
Growth by month
| Month (2026) | Transcripts published |
|---|---|
| January (from the 29th) | 6 |
| February | 1,122 |
| March | 1,766 |
| April | 1,767 |
| May | 2,462 |
| June | 2,287 |
| July | 3,922 |
| August | 4,317 |
| September | 5,533 |
| October (1 to 24) | 4,670 |
September had almost five times as many new transcripts as February. If October’s last seven days keep the pace of its first 24, the month will end at about 6,000, ahead of September.
What comes with each transcript
| Addition | Transcripts | Share |
|---|---|---|
| Detailed summary and key takeaways | 26,712 | 95.9 % |
| Questions and answers | 27,671 | 99.4 % |
| Chapters with timestamps | 17,502 | 62.8 % |
Chapter coverage is lower because chapters were added later and are still being filled in for older pages.
How we counted
- Scope. Every post of the transcript type with the status “published” on sozai.app at 05:13 UTC on 24 October 2026. That is 27,852 posts, and 51 drafts are not counted. We read the database with read-only queries and changed nothing.
- Language is the language of the caption track a transcript was built from, as recorded when the page was created. For Hebrew, the stored code is the legacy “iw”.
- Words is the word count saved with each transcript. Thirty-three transcripts have a count of zero.
- Video length is the duration reported by YouTube’s Data API for 25,863 videos. For the other 1,988, we used the end time of the last caption line instead. Where we have both, the median ratio between them is 1.00.
- Topics come from the topic labels assigned on 30 July 2026. Transcripts published after that date have no topic yet.
- Channels are the channel names YouTube reports for each video, counted as exact names.
- Months follow the publication date in the site’s time zone (London).
- Chapters counts only transcripts with at least one chapter. An empty chapter list does not count.
- We read no personal data: no IP addresses, no device identifiers, no account information.
