RankedSpeakFiller Words, Pace and Scores: What Our Practice Speeches Show (2026)
Last updated October 23, 2026 · about 900 speeches · nearly 200 speakers · April 6, 2026 to October 21, 2026
RankedSpeak is an iPhone app where you answer a prompt out loud for 30 to 90 seconds and get scored. We looked at every scored practice speech in the app, about 11 hours of speech in all, to see how people actually talk when they practise. These are aggregate numbers only. No recording, transcript or person is shown or can be picked out, and every figure describes at least 20 speakers.
Key findings
- The typical speaker says "um", "uh" or "you know" about 0.8 times a minute. Half of speakers fall between 0 and 2.4 a minute, and 37% never used one.
- Counting context fillers too ("like", "so", "basically" used as filler), the median rises to about 3.5 a minute, so most filler words are not "um" and "uh".
- "Um" is the most common hesitation sound: 47% of all um/uh/you-know fillers, used by 41% of speakers.
- Substance (the quality of the points made) is the weakest skill: it is the lowest of five scores for nearly half of speakers (48%). Clarity scores highest and was nobody's weakest skill.
- In a group of about 40 regular users who did at least five speeches, the average score rose from 6.1 to 6.5 out of 10, and hesitation fillers fell from 2.1 to 1.2 a minute. 64% scored higher on their fifth speech than their first.
- The median speaker talks at about 128 words a minute and uses 93% of the time they are given.
- About 62% of practice happens after 5 p.m. local time.
Sample: about 900 scored speeches by nearly 200 people, April 6, 2026 to October 21, 2026. Every speaker counts once. See methods for what that means and what the limits are.
How often people say um and uh
We counted hesitation sounds (um, uh, er, hmm) and "you know" used as a tic, straight from each transcript. Each speaker's rate is their total fillers divided by their total minutes, so someone who practised 30 times counts the same as someone who practised once.
| Measure | Per minute of speech |
|---|---|
| Median speaker: um, uh, you know | 0.8 |
| Middle half of speakers: um, uh, you know | 0 to 2.4 |
| All speech pooled: um, uh, you know | 1.1 |
| Median speaker: all fillers, including like, so, basically, actually, literally used as filler | about 3.5 |
| Middle half of speakers: all fillers | 1 to 7.7 |
| Hesitation filler | Share of all um/uh/you-know fillers | Share of speakers who used it |
|---|---|---|
| "um" | 47% | 41% |
| "uh" | 31% | 39% |
| "you know" | 23% | 20% |
The weakest speaking skill is substance
Every speech gets five scores from 0 to 10. We averaged each speaker's scores, then looked at which of the five was lowest for each person. Substance, whether the points themselves are specific and worth hearing, came out lowest most often. Clarity, whether you are easy to understand, was nobody's lowest score.
| Skill | Average score (0 to 10) | Share of speakers for whom it is the weakest skill |
|---|---|---|
| Clarity | 6.9 | 0% |
| Structure | 6.1 | 4% |
| Flow | 5.9 | 16% |
| Delivery | 5.9 | 32% |
| Substance | 5.6 | 48% |
What changes over the first five speeches
To compare like with like, this table follows one fixed group of about 40 regular users who recorded at least five speeches. Column by column it is the same people each time.
| Speech | Overall | Structure | Delivery | Substance | Um/uh/you know per min | Average length |
|---|---|---|---|---|---|---|
| 1 | 6.1 | 6.2 | 5.8 | 5.6 | 2.1 | 28 s |
| 2 | 6.4 | 6.5 | 6.0 | 6.2 | 1.8 | 43 s |
| 3 | 6.5 | 6.6 | 6.2 | 6.3 | 1.7 | 44 s |
| 4 | 6.4 | 6.5 | 6.1 | 6.2 | 1.3 | 49 s |
| 5 | 6.5 | 6.6 | 6.2 | 6.3 | 1.2 | 54 s |
Read this with care. There is no control group, the first speech is usually the short "introduce yourself" prompt every new user gets, and speeches got longer as people went on. People who stick with practice may also differ from people who stop. So the table shows that scores and filler rates moved in the right direction for this group, not that practice alone caused it.
Pace, time used and when people practise
| Measure | Result |
|---|---|
| Median speaking pace | 128 words per minute |
| Middle half of speakers | 111 to 149 words per minute |
| Speakers under 110 words per minute | 24% |
| Speakers over 160 words per minute | 17% |
| Share of the given time a typical speaker uses | 93% |
| Speeches that used less than half the time | 6% |
To turn these speeds into a time for your own speech, paste it into the speech length calculator.
| Local time of day | Share of practice |
|---|---|
| Morning (5:00 to 11:59) | 14% |
| Afternoon (12:00 to 16:59) | 23% |
| Evening (17:00 to 21:59) | 33% |
| Late night (22:00 to 4:59) | 29% |
| Prompt topic | Speakers (range) | Average overall score |
|---|---|---|
| Introduce yourself | over 150 | 6.0 |
| Relationships | 20 to 29 | 6.1 |
| Health | 20 to 29 | 6.2 |
| Everyday life | 20 to 29 | 6.3 |
| Education | 20 to 29 | 6.4 |
| Travel | 30 to 49 | 6.5 |
| Technology | 20 to 29 | 6.5 |
| Sports | 20 to 29 | 6.5 |
| Business | 20 to 29 | 6.6 |
| Music | 20 to 29 | 6.6 |
| Food | 30 to 49 | 6.7 |
Methods and limits
- Sample. Every scored practice speech recorded in the RankedSpeak iPhone app from April 6, 2026 to October 21, 2026 that was at least 10 seconds long and at least 20 words: about 900 speeches by nearly 200 people, about 11 hours in all, median length about 40 seconds. Judge Mode recordings (scoring a conversation between other people) are excluded.
- Every speaker counts once.We work out each person's average or rate first and then summarise across people, so heavy users (and any test accounts) cannot dominate a result.
- Transcripts and scores are AI-made. Each speech is transcribed and scored by an AI model (Google Gemini). The scoring rubric was adjusted during the period, so treat scores as a consistent yardstick within this data, not an absolute standard.
- Two filler counts.The "um, uh, you know" figures are counted directly from transcripts by a fixed pattern, so anyone re-running the query gets the same number. The "all fillers" figures are the app's own count, which also decides whether words like "like" and "so" were used as filler. Before 19 October 2026 that count came from the AI model and could be off by around 10%, so we round it.
- Who these people are. RankedSpeak users chose to practise speaking, mostly in English, on a phone, answering short impromptu prompts. They are not a random sample of the population, and a 30 to 90 second answer is not a talk or a meeting.
- Privacy. Only aggregate numbers are published, and sample sizes are rounded. No figure describes fewer than 20 speakers, and no recording, transcript, username or score of an individual is shown. See our privacy policy.
- Updates. We plan to re-run the same queries each quarter and update this page. The date at the top shows the last refresh.
Cite this
RankedSpeak (2026). "Filler Words, Pace and Scores: What Our Practice Speeches Show (2026)." rankedspeak.com/data/speaking-practice-statistics. Data from April 6, 2026 to October 21, 2026; accessed [date].
You are welcome to quote these numbers with a link to this page. The data is licensed CC BY 4.0. Questions about the data: support@rankedspeak.com.