Does Practice Actually Improve Speaking? What 12,660 Sessions Tell Us

"Practice makes perfect" is the kind of thing nobody argues with and almost nobody verifies. So we tried to verify it — for speaking, specifically, on our own users.

We pulled every scored speaking session from Elqo's production database and looked at how individual users' scores changed as they completed more sessions. 12,660 scored attempts. 1,033 users with at least three sessions. Every session graded on the same 0–100 scale by the same model. The question was simple: do scores actually go up with reps, or are we all just kidding ourselves?

Short answer: yes, they do — and the effect is larger and more consistent than I expected.

The Headline Numbers

Before getting into the methodology, here's what 1,033 users' practice histories actually say.

Metric	Value
Users with 3+ scored sessions	1,033
Users with an improving trend	666 (64%)
Users with a declining trend	354 (34%)
Average points gained per additional session	+1.9
Mean score at session 1	43.3
Mean score at session 10	55.9
Mean score at session 20	60.1
Mean score at session 30	62.6
Absolute lift, session 1 → 30	+19.3 points

Two out of three users improve. The average user picks up roughly two points per additional session early on. And in aggregate, scores climb from the low 40s to the low 60s over the first thirty sessions — a 45% relative improvement.

That last number is the one that gets you nodding. It's also the one a good skeptic would push back on hardest. So let's deal with the obvious objection first.

"But What About Survivor Bias?"

The aggregate average has a built-in problem: by session 30, only 46 of the original 1,033 users are still in the dataset. The other 987 dropped out somewhere along the way. It's entirely possible that the people who keep practicing are the people who were good (or improving) to begin with — and that the score lift is really just self-selection, not practice.

So we re-ran the analysis with the same users tracked from start to finish. For each cohort below, we took only the users who reached at least K sessions, and compared their average score at session 1 to their average score at session K. Same exact people at both ends.

Cohort	Users	Session 1 avg	Session K avg	Lift
Reached 5+ sessions	445	45.96	52.04 (sess 5)	+6.1
Reached 10+ sessions	171	48.64	55.91 (sess 10)	+7.3
Reached 20+ sessions	73	49.48	60.05 (sess 20)	+10.6

The 73 users who hit at least 20 sessions improved by 10.6 points from their first session to their twentieth. That's not survivor bias talking — it's the same group of people, measured at both ends of their own practice arc. The improvement is real.

You can watch it happen, session by session, in that 20+ cohort:

Session #	Avg score	Session #	Avg score
1	49.48	11	58.77
2	50.99	12	59.66
3	53.32	13	57.99
4	55.40	14	61.71
5	58.79	15	62.05
6	57.96	16	57.95
7	56.62	17	60.75
8	59.21	18	60.03
9	59.19	19	60.23
10	59.74	20	60.05

Two things to notice. The trajectory is bumpy but unmistakably upward. And most of the lift comes in the first five sessions — from 49.5 to 58.8. After that, it's smaller, slower gains layered on top.

Improvement Holds Across Every Engagement Level

The next worry was that improvement might be a "casual user" phenomenon — people start low, get a few quick wins, then plateau. So we grouped users by their total session count and compared the mean of the first half of each user's sessions to the mean of their second half.

Total sessions	Users	First half avg	Second half avg	Δ	% improved
3–4	588	41.65	45.75	+4.11	60%
5–9	274	46.83	50.52	+3.69	62%
10–19	98	49.51	53.16	+3.64	63%
20–49	52	56.88	60.52	+3.64	69%
50+	21	61.16	64.13	+2.97	71%

Two clean findings from this table:

1. Every engagement bucket improves. The second half is always meaningfully higher than the first. There's no "ceiling" cohort. Even users with 50+ sessions are still gaining ~3 points between their early and late attempts.
2. The share of users who improve rises with engagement. 60% of casual users (3–4 sessions) improve. By the 50+ tier, it's 71%. Sticking with practice doesn't just produce more reps — it raises the probability that any given user is on an upward trajectory at all.

Yes, heavier users also start higher. A user who eventually completes 50+ sessions begins around 61, vs ~42 for someone who only does 3–4. So there's clearly some self-selection — people who are already decent stick around longer. But the improvement is happening inside each group, regardless of where they started.

Diminishing Returns Are Real (And That's Fine)

The "+1.9 points per session" headline hides an important detail: the per-session slope shrinks fast.

Total sessions bucket	Avg slope (points per session)
3–4	+2.65
5–9	+1.26
10–19	+0.47
20–49	+0.24
50+	+0.10

In other words: most of the gain happens in the first 10–15 sessions. After about session 20, scores stabilise in the low-to-mid 60s and creep up slowly from there. That's exactly what you'd expect from a skill-acquisition curve — rapid early learning, then a long, slow grind for the marginal improvements.

This is good news, not bad news. It means you don't need to practise forever to see meaningful improvement. You need to practise about 10–15 times. Two or three short sessions a week and you're inside the steepest part of the curve within a month.

What This Actually Means For You

If you take one thing from this analysis, take this: the first 10 sessions matter more than the next 90. The biggest, fastest gains in our data come from going from "I have never recorded myself" to "I have recorded myself ten times." If you're on the fence about whether practising is worth your time, the data says you can answer that question for yourself in about 10 reps.

A few specific implications:

1. Treat the first five sessions as a learning sprint. Don't agonise over each one. The cohort data shows the average user adds 9 points between session 1 and session 5 — almost as much as the next 15 sessions combined. Volume beats perfectionism here.
2. Aim for a minimum of 10 sessions before you judge yourself. 64% of users have an improving trend overall, but the early-session noise is real. Three sessions isn't enough to know whether you are improving. Ten usually is.
3. If you've already done 20+ sessions, switch your goal. The slope at this point is ~0.2 points per session. Chasing the score is going to feel slow. The leverage is in targeted drills on your two or three specific weak spots, not in raw volume.

A Note On The Method

For the data nerds: every "session" here is one row in lesson_attempts — one full scored recording at a lesson. Scores come from the same 0–100 model applied uniformly across the period. We didn't use the progress table (which only stores the latest score per lesson) or lesson_completion_events (no score column), because neither lets you reconstruct a per-session trajectory. The 1,033-user cohort is everyone with at least three scored attempts, which is the minimum to fit a meaningful per-user trend line.

The one honest confound we can't fully resolve is engagement vs ability: users who do 50+ sessions also start about 20 points higher than users who only do 3–4. So "more practice causes more improvement" is partly tangled up with "better speakers tend to practise more." But the within-cohort improvement is real and significant at every level — including for the users who started below average.

The Bottom Line

"Practice makes perfect" is wrong on a technicality: nobody in our dataset is perfect, and the curve flattens long before anyone gets close. But "practice produces measurable, repeated, statistically-real improvement in two out of three people who try" is what the data actually says, and that's a much more useful claim.

If you've been quietly assuming you're the exception — that you're someone who plateaus, or that AI feedback wouldn't move the needle for you — the numbers say there's a 64% chance you're wrong. The cheapest way to find out is to spend the next 10 sessions on it and check.