Across 10,446 YouTube videos in the SozAI transcript library, English speakers said a median of 168 words per minute of speech, and half of the English videos fell between 147 and 192. Spanish and Russian both came in at 139, French at 165, German at 130 and Portuguese at 149. We measured 14 languages from the timestamps of each video’s own captions. We left out music and adverts, and also dubbed tracks and every transcript whose caption text did not hold up.
Measured 31 October 2026. Method version 1.1. 10,446 videos, about 6,880 hours of speech. Source: the public transcript library on sozai.app.
Quick facts
| Measure | Value |
|---|---|
| Languages published | 14, each with at least 100 videos |
| Videos measured | 10,446 |
| Speech measured | about 6,880 hours |
| English, median | 168 words per minute |
| English, middle half of videos | 147 to 192 words per minute |
| English talks and speeches, median | 147 words per minute (56 videos) |
| Checked against YouTube’s own captions | 106 videos, median word-count ratio 0.99 |
Speaking rate by language
Each video gives one rate: words divided by minutes of speech. The median is the middle video of the language. The middle half is the range from the 25th to the 75th percentile, so half of the videos fall inside it. The last column is the 95 % bootstrap interval of the median, which shows how much the median itself could move with a different sample of the same size.
| Language | Videos | Median, words per minute | Middle half | 95 % interval of the median |
|---|---|---|---|---|
| English | 3,692 | 168 | 147 to 192 | 167 to 170 |
| Russian | 2,884 | 139 | 122 to 155 | 138 to 140 |
| Polish | 634 | 128 | 118 to 140 | 127 to 130 |
| Arabic | 612 | 123 | 104 to 143 | 120 to 125 |
| Hindi | 430 | 181 | 158 to 207 | 177 to 188 |
| Spanish | 382 | 139 | 121 to 164 | 135 to 143 |
| French | 377 | 165 | 137 to 203 | 161 to 168 |
| Ukrainian | 349 | 121 | 103 to 135 | 118 to 124 |
| Indonesian | 347 | 129 | 115 to 145 | 127 to 133 |
| Korean | 200 | 112 | 97 to 127 | 108 to 115 |
| German | 166 | 130 | 107 to 170 | 124 to 137 |
| Portuguese | 160 | 149 | 135 to 170 | 145 to 155 |
| Hebrew | 107 | 138 | 120 to 156 | 131 to 144 |
| Bengali | 106 | 144 | 128 to 168 | 139 to 152 |
Compare the numbers within a language, not across languages. A word here is whatever stands between two spaces, and languages pack speech into words differently. Korean writes a noun and its particle as one word, so its count is low (112). Hindi writes many short grammatical words separately, so its count is high (181). Neither number says that Korean speakers talk slower or Hindi speakers faster than anyone else.
Languages we left out
A language appears only when at least 100 of its videos pass every check. These did not: Japanese 95, Italian 93, Chinese 87, Turkish 33, Vietnamese 32, Dutch 23, Thai 23, Swedish 22, Persian 22. Languages with fewer than 20 measured videos are not listed. Japanese, Chinese and Thai are written without spaces between words, so for them we count characters per minute instead of words. Many Japanese and Chinese transcripts failed the text checks described below, which is why so few remained.
By type of video
We sorted videos into types by words in the title and the channel name: interview, podcast, lecture, lesson, TED, keynote, “how to” and their equivalents in the other languages. 1,612 of the 10,446 videos matched a type. In a spot check of 100 sorted titles, about 9 in 10 were in the right type. To compare types across languages, each video’s rate is divided by the median of its own language, so 1.07 means 7 % faster than usual for that language.
| Type | Videos | Compared with the language median | Middle half |
|---|---|---|---|
| How-to and advice | 442 | 7 % faster (1.07) | 0.89 to 1.23 |
| Lecture, lesson or course | 421 | 2 % slower (0.98) | 0.81 to 1.13 |
| Podcast | 187 | 4 % faster (1.04) | 0.89 to 1.16 |
| Talk or speech | 125 | 6 % slower (0.94) | 0.83 to 1.07 |
| Interview | 109 | 1 % slower (0.99) | 0.93 to 1.11 |
The differences between types are small next to the spread inside each type. Talks and speeches are the slowest. How-to and advice videos are the fastest, and many of them are scripted and edited tightly. Five more types had fewer than 100 videos each and are not shown: news, audiobooks, sermons, gaming and drama.
English by type
| Type | Videos | Median, words per minute | Middle half |
|---|---|---|---|
| How-to and advice | 282 | 184 | 159 to 211 |
| Podcast | 69 | 178 | 156 to 192 |
| Lecture, lesson or course | 92 | 177 | 147 to 203 |
| Interview | 55 | 166 | 143 to 187 |
| Talk or speech | 56 | 147 | 131 to 176 |
English talks and speeches ran at a median of 147 words per minute, 21 below the median of English YouTube as a whole. If you are timing a speech you will give to an audience, that is the closer number. In Russian, lectures and lessons ran at 124 words per minute (131 videos), against a Russian median of 139.
By topic
Transcripts created up to 30 July 2026 carry one of 20 topic labels, and 3,860 of the measured videos have one. The scale is the same as for types. Topics with fewer than 100 videos are not shown.
| Topic | Videos | Compared with the language median | Middle half |
|---|---|---|---|
| Science & Education | 735 | 7 % slower (0.93) | 0.80 to 1.02 |
| Business & Startups | 430 | 9 % faster (1.09) | 0.98 to 1.22 |
| Psychology & Self-Development | 375 | 4 % slower (0.96) | 0.86 to 1.07 |
| Tech & AI | 323 | 1 % faster (1.01) | 0.89 to 1.15 |
| Software How-To | 243 | 1 % faster (1.01) | 0.87 to 1.19 |
| Money & Investing | 236 | 9 % faster (1.09) | 0.96 to 1.16 |
| Health & Fitness | 235 | 2 % slower (0.98) | 0.87 to 1.11 |
| News & Politics | 224 | 1 % slower (0.99) | 0.91 to 1.08 |
| Religion & Spirituality | 172 | 7 % slower (0.93) | 0.83 to 1.04 |
| Gaming | 148 | 2 % slower (0.98) | 0.83 to 1.12 |
| Society & Culture | 117 | 1 % slower (0.99) | 0.89 to 1.12 |
| Film & TV | 109 | 5 % slower (0.95) | 0.78 to 1.13 |
| Lifestyle | 100 | 2 % faster (1.02) | 0.87 to 1.15 |
Business and money videos ran 9 % faster than their language’s median. Science and education, and religion, ran 7 % slower.
How we measured
- Source. Every published transcript on sozai.app at 10:55 UTC on 31 October 2026: 29,514 posts. Each was built from the captions YouTube shows for the video and keeps their timestamps. We read the database with read-only queries and read no personal data.
- Language is the language of the caption track the transcript was built from.
- Words are the units between spaces that contain a letter or a digit, after sound tags such as [Music] and [Applause] are removed. For Japanese, Chinese and Thai we count letters and digits instead.
- Speech time. Captions are stored as segments of about 40 words, each with a start and an end time. Speech time is the time those segments cover. A segment that runs at less than half the usual pace of its own video holds a long pause, music or a cut, so we drop its words and its time together.
- Rate is words divided by speech time, one number per video. The tables give the median over videos, so a three-hour lecture counts as much as a five-minute clip.
- Minimums. At least 2 minutes of speech per video, 100 videos for a language, 100 for a type or a topic, and 50 for one type within one language.
What we left out
| Reason | Transcripts |
|---|---|
| Not built from timed captions: older transcripts whose timestamps a model estimated, or no source recorded | 5,176 |
| Music, songs, recitation, ASMR (topic label, title words or [Music] tags in the captions) | 4,465 |
| Created after 20 October 2026, 13:56 UTC (see the limits below) | 2,274 |
| The punctuation step shifted or repeated lines (a segment of more than 100 words) | 2,262 |
| Shorter than 2 minutes, as a video or in measured speech | 1,447 |
| Caption language does not match the title: probably a dubbed or translated track | 1,070 |
| Missing caption cues (more than 15 % of segments hold fewer than 15 words) | 941 |
| Incomplete records: no title or language, or fewer than 5 segments | 508 |
| Japanese, Chinese and Thai text shortened by the punctuation step | 227 |
| Silence- or music-heavy (more than 30 % of caption time at under half the video’s pace) | 85 |
| Trailers and adverts | 30 |
| Outliers (more than three interquartile ranges outside the middle half of the language) | 12 |
| Left out in total | 18,497 |
That leaves 11,017 measured videos, of which 10,446 are in the 14 published languages. Two settings decide the most: the pause floor (half the video’s usual pace) and the outlier fence. Moving the floor to 0.4 or 0.6 changes no language median by more than 2 words per minute, and dropping the outlier fence changes none.
Checked against YouTube
On 31 October 2026 we picked 112 measured videos at random, 8 per language, and downloaded YouTube’s own speech-recognition captions for each. 6 videos no longer had that track. For the other 106:
| Check | Result |
|---|---|
| Our word count divided by YouTube’s, median | 0.99 |
| Videos within 10 % of YouTube’s word count | 92 % |
| Videos whose rate is within 10 % of the rate from YouTube’s track | 97 of 106 |
| Median difference in rate | −0.4 % |
The rules above were fixed before this check. An earlier check on a different sample is how we found two of them: videos with many short caption segments had lost a third or more of their words, and segments of more than 100 words held repeated text.
Limits
- These are YouTube videos. Many are scripted and edited, and editing cuts pauses, so they run faster than unscripted speech in a room.
- The sample is what visitors asked us to transcribe, not a random draw from YouTube. English and Russian make up 63 % of the measured videos.
- Automatic captions miss words in noise and crosstalk, and every missed word lowers the rate a little.
- We cannot tell automatic captions from captions a person wrote, and both are in the sample.
- Many videos have more than one speaker. The rate belongs to the video, not to one person.
- Transcripts created from 20 October 2026, 13:56 UTC, on are not in this edition. That day the step that adds punctuation to captions changed, and in many of the newer transcripts it repeated or dropped lines, so we left the whole period out.
Use the numbers
The speech time calculator and the teleprompter start from the median of your page’s language where we have one, and you can change the rate. The word counter uses the same median for its speaking time.
