# State of Speaking Practice, November 2026

ISSEN, November 2026. This page and its Markdown copy use the same figures.

Cite this report: ISSEN, State of Speaking Practice, November 2026. https://www.issen.com/state-of-speaking-practice/

## Headlines

From 20 August through 18 November 2026, 124,397 people spent 2,135,813 minutes in 337,498 conversations. The median conversation lasted 4.3 minutes, and the average lasted 6.3.

German accounts for 42% of those minutes. English accounts for 35% of the people and 27% of the minutes.

The largest path is Arabic speakers practicing German: 26,173 people and 32% of all minutes in this window.

A real conversation is the first session of at least 5 minutes. Of 82,092 new learners, 50.3% reached one within 28 days. Among those who did, the median was the same day, in the first session.

In the learner's own timezone, the busiest hour is 9 pm. Sessions that start around 6 to 8 am are fewer and longer.

The median session is 5.6 minutes on the web, 4.6 minutes on iOS, and 3.8 minutes on Android.

Swedish sessions are the longest on average, at 10.6 minutes, among languages with at least 30 learners. Among the large languages, Portuguese sessions are the shortest, with a median of 1.7 minutes.

In a sample of 60 Korean sessions of at least 5 minutes, 39 of the longest learner turns were in another language.

## Languages

Shares of sessions and minutes are shares of all practice in the window. Shares of people can add up to more than 100% because one person can practice more than one language. Languages with fewer than 30 people are left out of the ranked list. Forty-one languages cleared that line.

| Language | People | Sessions | Minutes | Share of minutes | Median minutes |
| --- | --- | --- | --- | --- | --- |
| German | 35,532 | 115,619 | 899,885 | 42.1% | 5.2 |
| English | 43,438 | 100,894 | 585,276 | 27.4% | 4.3 |
| French | 7,859 | 24,104 | 162,617 | 7.6% | 4.8 |
| Spanish | 8,992 | 23,398 | 139,266 | 6.5% | 4.1 |
| Italian | 4,518 | 11,236 | 65,525 | 3.1% | 4.0 |
| Portuguese | 8,571 | 15,387 | 42,595 | 2.0% | 1.7 |
| Japanese | 4,365 | 9,215 | 39,780 | 1.9% | 3.1 |
| Dutch | 1,336 | 4,619 | 30,025 | 1.4% | 3.7 |
| Hebrew | 826 | 2,713 | 21,637 | 1.0% | 4.7 |
| Greek | 750 | 2,431 | 19,347 | 0.9% | 5.5 |
| Korean | 2,696 | 5,479 | 18,759 | 0.9% | 2.3 |
| Swedish | 415 | 1,342 | 14,223 | 0.7% | 5.5 |

Swedish has the highest average, 10.6 minutes. Hebrew averages 8.0, Greek 8.0, and German 7.8. Portuguese averages 2.8.

## Source language and target language

The source language is the language the learner set as one they already speak. It is the closest field the database has to a native language.

| Source | Target | People | Minutes |
| --- | --- | --- | --- |
| Arabic | German | 26,173 | 682,241 |
| Arabic | English | 12,249 | 195,021 |
| Spanish | English | 11,819 | 203,767 |
| Portuguese | English | 8,533 | 46,957 |
| English | English | 4,760 | 43,107 |
| Portuguese | Portuguese | 4,557 | 15,872 |
| English | German | 3,357 | 96,016 |
| English | Spanish | 2,833 | 75,332 |
| English | French | 2,662 | 70,923 |
| German | German | 2,706 | 45,173 |

Pairs with fewer than 30 people are omitted. Some people practice the same language they set as their source language. That is common for English and Portuguese.

## How long a session is, and how often

A session counts when its stored duration is greater than zero. Duration is the length of the call in seconds, including the tutor's speech and pauses. The median call is 4.3 minutes. A quarter are under 1.4 minutes, and a quarter are over 6.1. Nineteen percent are under one minute. 44.8% last at least 5 minutes.

In a week when someone practiced, the median was 1 session and 5.6 minutes. The average was 2.2 sessions and 13.7 minutes. The median person's longest run of consecutive practice days in this window was 1 day. 740 people reached a run of 7 days, 252 reached 14, and 78 reached 30.

## Time to a real conversation

A real conversation is the learner's first session of at least 5 minutes. That bar is a real back-and-forth: in 63,798 first 5-minute sessions that had a stored transcript, the median learner took 17 turns, and 98.6% took at least 6. Whitespace word counts are a poor measure for Japanese, Korean, and Chinese, so the turn count is the check.

The new-learner group is 82,092 people whose first session ever, of any length above zero, fell between 20 August and 21 October 2026, so each person had 28 days to be observed. 50.3% reached a 5-minute session in those 28 days. 36,759 of them did it in their first session. Among the people who reached it, the median time was the same day and the median number of sessions was 1.

A split by language did not finish within the read-only query limit. The published figure is the overall one.

## When people practice

The clock is the learner's timezone. Every session in this window had a timezone name Postgres accepts. Ten learners were set to UTC.

The week is almost flat. Wednesday has the most minutes, 323,835. A weekday averages about 311,000 minutes, and a weekend day about 289,000. The busiest clock hour is 9 pm, with 24,178 sessions. The longest median sessions start around 6 am (5.5 minutes), 7 am (5.2), and 8 am (5.3). Those morning hours have far fewer people.

## Platform

| Platform | People | Sessions | Minutes | Median minutes |
| --- | --- | --- | --- | --- |
| Android | 72,798 | 182,146 | 1,073,413 | 3.8 |
| iOS | 47,074 | 140,245 | 923,505 | 4.6 |
| Web | 5,295 | 15,107 | 138,895 | 5.6 |

A person can appear on more than one platform. Web sessions are the longest. Android has the most people and the most minutes.

## Common difficulties, by language

Tutor corrections are generated during the call and sent to the app. They are not stored in a table. There is no stored list of grammar categories, corrected words, or pronunciation scores.

Two reads of the same window are used instead.

The first is a sample of transcripts. For each of the eight languages with the most learners, up to 60 different people were taken from sessions of at least 5 minutes between 20 October and 18 November 2026. German and Japanese had 58 people in that slice. The longest learner turn in the session was read and labeled. A session can also be marked when another writing system appears anywhere in the learner's turns. The counts below are that sample. They are a floor: a shorter turn in the same session can contain an error the longest turn does not. Examples are rewritten patterns, not quotations.

The second is a judge that has already labeled whole sessions. Its "mishearing" flag means the transcript is garbled or the tutor reacts to a mis-hearing. Its "correction" flag means the tutor corrected grammar, vocabulary, or pronunciation. Those percentages use every judged session in the full window, which is a much larger base than the 60-session sample.

### German, 58 sampled sessions

Another writing system appears in 37 of 58 sessions, usually Arabic. On the longest turn, articles, gender, or case are off in 10 sessions, word order in 4, and a preposition in 3. Rewritten patterns: a feminine article on "bicycle", and a "because" clause with the verb in the middle.

The judge marked a correction in 76.4% of 87,648 rated German sessions, and mishearing in 13.6%.

### English, 60 sampled sessions

The longest turn is mainly another language in 7 sessions. The verb form is off in 5, word order in 3, articles in 3, and a preposition in 3. Rewritten patterns: "I will to play", and "the jacket blue".

The judge marked a correction in 71.3% of 72,521 rated English sessions, and mishearing in 16.7%.

### Spanish, 60 sampled sessions

The longest turn is mainly another language in 19 sessions. Gender or article is off in 4, the verb form in 3, and a preposition in 3. Rewritten patterns: a masculine article on "situation", and "it depends in the seats".

The judge marked a correction in 69.9% of 16,789 rated Spanish sessions, and mishearing in 18.0%.

### Portuguese, 60 sampled sessions

The longest turn is mainly another language, often Spanish, in 9 sessions. An infinitive standing in for a conjugated verb appears in 2, and the house takes the masculine contraction in 1. Many Portuguese sessions are people whose source language is also Portuguese, so learner-grammar errors are scarce in this sample. Rewritten pattern: "I to eat at home".

The judge marked a correction in 43.7% of 8,075 rated Portuguese sessions, and mishearing in 33.4%. Mishearing is the strongest Portuguese signal in the stored labels.

### French, 60 sampled sessions

The longest turn is mainly another language in 15 sessions. A preposition is off in 4, and a verb form in 3. One session uses a masculine partitive with "mustard". Rewritten pattern: "an interest in languages" with the preposition that means "inside".

The judge marked a correction in 74.7% of 17,982 rated French sessions, and mishearing in 19.8%.

### Italian, 60 sampled sessions

The longest turn is mainly another language in 16 sessions. An article or gender is off in 3, a preposition in 2, and a verb form in 2. Rewritten pattern: "because Italy is a beautiful country" with the article left out.

The judge marked a correction in 71.3% of 7,952 rated Italian sessions, and mishearing in 18.5%.

### Japanese, 58 sampled sessions

The longest turn is mainly another language in 10 sessions. A particle is off in 4, and a verb form in 2. Rewritten pattern: "I work at a company" with the particle に where で belongs.

The judge marked a correction in 56.0% of 6,070 rated Japanese sessions, and mishearing in 26.7%.

### Korean, 60 sampled sessions

The longest turn is in another language in 39 sessions, and too short or mixed to treat as a Korean sentence in 19. Only 2 longest turns are Korean, which is too few to rank grammar categories. One learner asks which polite ending matches the vowel. That is a single case, not a rate.

The judge marked a correction in 61.8% of 3,220 rated Korean sessions, and mishearing in 39.5%.

## Methodology

Window: sessions that started from 20 August 2026 00:00 UTC through 18 November 2026, excluding any session that started in the last hour before the query. Duration is greater than zero. Usernames matching test, demo, load-test, autosana, or a tester domain are excluded. A cell with fewer than 30 people is dropped.

The database read was a read-only analytics role. Nothing was written.

What this report does not include: a per-language time to the first 5-minute session, a stored correction category, a pronunciation score, and any payment or renewal figure.
