How accurate is Akousa?
Last updated 6 October 2026
You read one short paragraph aloud, and Akousa checks every sound of every word on your iPhone against native American or British speakers, then shows you the words that slipped.
This page shows what we have measured, how we measured it, and what we don’t know yet. Every number comes from our own test records. Akousa’s marks are automatic, so they can be wrong.
of sounds land in the native range for American speakers our lines never saw. British: 94.6%.
of 77 words: what native speakers get on our paragraph in the default goal, estimated from their recordings. American alone: 59 to 71. The app shows the lower end as the bar to reach.
how closely Akousa ranks 250 learners the way five experts do. 1 would be the same order.
How native speakers score
Akousa calls a sound native when it lands where native speakers put it. To learn where that is, we measured recordings of real native speakers, never computer voices:
- American English: 80 speakers reading audiobooks (LibriSpeech), 377,276 sounds.
- British English: 33 speakers from England, recorded in one studio (VCTK), 57,752 sounds.
For each sound we draw a line that 95 in 100 native sounds clear. Then we test the lines on native speakers they never saw:
| Native speakers the lines never saw | Sounds in range |
|---|---|
| American, 23 speakers (LibriSpeech) | 95.2% |
| British, 12 speakers from England (VCTK) | 94.6% |
| American, 4 speakers from another recording set (CMU ARCTIC)2 Oct | 95.8% to 96.5% |
| Canadian, 1 speaker (CMU ARCTIC), judged as American2 Oct | 95.8% |
So even native speakers have a few sounds in every 100 outside the range. Everyone’s speech varies, and that is fine. Across all 80 American speakers, including those who set the lines, each had 93% to 98% of their sounds in range; across the 33 British speakers, 92% to 97%.
On our paragraph
Your result says how many of the paragraph’s 77 words sound native, meaning every sound in the word lands in the native range. Under your number, Akousa shows the bar for the goal you chose: the number 95 in 100 native readings reach. For “Be clearly understood” it reads “Native speakers: 62 or more”; for American, 59; for British, 57. Here is the whole range of what native speakers get, so you can see that they miss a few words too:
“Be clearly understood” counts a word when it lands in either American or British English, so native speakers clear a few more. Even native speakers rarely get all 77.
No native speaker has read this paragraph for us yet. So we built 3,000 readings of it from native speakers’ own sounds in other sentences, using only speakers the lines never saw, and show where the middle 90 in 100 land.
Read the paragraph
In my first week in Vermont, I asked a waiter for the dessert menu. He smiled and said the desert was three thousand miles west. On Wednesday, I offered thirteen dollars for a small basket of strawberries, and the farmer heard thirty, with visible pleasure. Nobody laughed at me; we laughed together. Walking home that evening, I realized something: clear speech is not about losing your voice. It is about being heard the way you mean it.
Catching slips
Accuracy has two sides. When a sound slips, does Akousa notice? And when a sound is fine, does it leave it alone?
What the marks mean
- west
- Red line. A listener may hear another sound. Only for a sound below a line that 99 in 100 native sounds clear.
- evening
- Grey dots. Close, not native yet. Outside the usual native range, but not that far out.
- home
- No mark. The word sounds native.
Against expert ears
speechocean762 is a public research set: 5,000 English sentences read by 250 people whose first language is Mandarin, children and adults. Five experts scored every sound. Our lines come only from native speakers, so this is an outside check.
- Person by person. Line up the 250 speakers from strongest to weakest. Akousa’s order agrees closely with the experts’ order: +0.86 for the share of sounds in range, and +0.83 for the share of words that sound native. On this scale, 1 is the same order and 0 is no link.
- Sound by sound. Of the sounds the experts scored lowest, Akousa marked 97 in 100. Of the sounds they scored highest, it marked 38 in 100 as not yet native. In between, the lower the experts’ score, the more often Akousa marked the sound.2 Oct
The 38 is by design. Expert raters are more forgiving than our native lines: Akousa asks whether a sound lands where native speakers put it. That is why there are two marks, and red is kept for sounds a listener may hear as another sound.
Word stress
Akousa also checks which part of a longer word you stress, and mentions it only when it is clearly off. On the same expert set, it marked 21.9% of the words the experts heard as wrongly stressed, and 1.4% of the words they heard as well stressed. Native speakers get 0% to 1.6% of their words marked. So it misses many stress slips, but rarely marks a good one.
Computer voices
We also plant slips in computer voices, where we know every answer. On our current paragraph, a clean reading gets none of the six practice sounds marked (V, W, TH, Z, F and an extra vowel before S), and readings with planted slips get exactly the planted sounds, on the phone and on our reference scorer alike. A voice is not a person, so this is only a sanity check.
Real readings of our paragraph
53 people have read our paragraph for our accent study so far: 20 whose first language is Arabic and 20 whose first language is Tagalog, through Prolific, and 13 who grew up speaking Hindi, Telugu, Tamil, Malayalam, Bengali or Urdu, by a direct link. Every recording has been checked for being clear and complete, and scored. Counting the slips Akousa catches in them needs a listener to mark every sound by ear first. That hasn’t been done yet, apart from the one check below.
One sound checked by ear
For Arabic speakers we are adding a check of whether P comes out like B, as in “pleasure” and “speech”. It isn’t in the app yet. One person on our team listened to 30 of their study clips without seeing Akousa’s verdicts: 15 the check marked as B and 15 it left alone. His ear agreed with the check on 23 of 30 (77%). Of the 15 marked, he heard B in 11; of the 15 left alone, he heard B in 3. He heard all 4 native speakers’ clips as P, as the check did.5 Oct This is one listener, who is not a native English speaker, on one sound. It is a first check, not a measured catch rate; native English listeners are the stronger test, and that study is still to come.
New checks are tried on native speakers first
- Trap words, for the 21-day plan. Everyday words that are easy to say the way they are spelt, such as receipt, salmon, debt, listen and Wednesday. A trap word ships only if natives’ own recordings set it off in at most 1 in 100 of their words with the same pattern, a rule written before the first test. Of 64 words checked, 12 passed. For example, “of” set it off in 1 of 5,492 native words, from 187 speakers, and the silent b of comb, climb, thumb and lamb in 2 of 4,010 native words that end in M. Five were dropped because natives tripped them, among them Christmas and usually.6 Oct
- The “-ed” ending, being tested. Some readers say “laughed” with a voiced ending, close to “lavd”. We are testing a check for the ending that looks only where native speakers don’t blur it themselves. On native speakers’ recordings it flagged 10 of 2,388 “-ed” words (0.4%). It isn’t in the app yet.6 Oct
Across first languages and accents
Everyone is measured against the same native lines. A lower share here doesn’t mean worse English. It shows where an accent differs from the target, which is what Akousa is built to find.
| Judged as American English | Sounds in range |
|---|---|
| American, 23 speakers the lines never saw (LibriSpeech) | 95.2% |
| Scottish, 1 speaker (CMU ARCTIC)2 Oct | 94.4% |
| Indian English, 1 speaker (CMU ARCTIC)2 Oct | 85.8% |
| Mandarin speakers learning English, 250 (speechocean762)Most between 30% and 78% | 60% |
| South Asian English, 2,453 speakers (Common Voice)6 OctHalf of them between 71% and 86% | 78.2% |
| Judged as British English | Sounds in range |
|---|---|
| From England, 12 speakers the lines never saw (VCTK) | 94.6% |
| American, 20 speakers (VCTK) | 93.6% |
| Scottish, 19 speakers (VCTK) | 93.1% |
| Indian English, 2 speakers (VCTK) | 89.7% |
The marks land where these accents really differ. For the Indian English speaker judged as American, most fell on R (rolled or tapped, 66 times in 100 sentences), W said like V (15) and Z said like S (9). For the Scottish speaker, on the rolled R (38).
Women and men, and age
We ran the scorer on 10,125 short recordings by 2,453 South Asian English speakers in Common Voice, Mozilla’s public collection of read sentences, at most 5 each. Of those who said, 247 are women and 1,795 men.6 Oct
- Women and men. The six sounds slipped in 9.4% of the words checked for women and 9.5% for men: the same. Native speakers on the same checks: 1.2% and 1.1%.
- Age. Across the six sounds, no age group differs from people in their twenties: teens, thirties and 40 and over alike.
- Every sound. On the score your result is built on, the women and older speakers in this set land a little closer to native speakers: 2.2 points more of women’s sounds in the native range than men’s, and 3.7 more for people 40 and over than for people in their twenties. Native women and men score alike (95.0% and 95.1%), so the lines themselves treat them the same.
These are the scorer’s verdicts, not a listener’s.
A lean toward American, found and fixed
The speech model we build on learned from American dictionary pronunciations. At first it heard an English speaker’s T between vowels, as in “thirty”, as the American soft T, and marked it 14.1% of the time (T in other places: 1.2%). British speakers now have their own line there: their T between vowels went from 87.6% to 97.4% in range. Indian English speakers are still marked on T and R there 2.8 times as often as native speakers, so real differences still show.
Which English you sound closest to
We are testing one more line for the result: the English you sound closest to. It isn’t in the app yet. It is built to say nothing rather than guess wrong:
- Indian speakers of 19 first languages (Svarah, 217 readings): “South Asian English” for 157 readings (72%), nothing for 60, a wrong family for none.
- European speakers of 15 first languages it has no group for (VoxPopuli, 451 readings by 262 people): a family for none.
- Mandarin speakers (speechocean762, 250): a wrong family for 2.
- Women and men (Common Voice, 2,453 South Asian English speakers): it named “South Asian English” for 61% of the women and 70% of the men. When it is unsure it says nothing, and it almost never names a wrong accent (0.4% and 0.6%). So we checked whether it is less sure of women in general. In four other sets where gender is known, the gap flips: Indian English speakers recorded on their own devices, women named 18 points more often than men; in a studio, 4.5 points less; native speakers, no gap. So this gap comes from these recordings, not from the recogniser, and it was always silence, never a wrong accent.6 Oct
We don’t guess your mother tongue
We tested naming five Indian mother tongues from one reading. On new speakers it was right 26% of the time, no better than always guessing the most common one. So Akousa doesn’t try.
What we don’t know yet
- Whether native listeners agree. We haven’t yet run a study where native American and British listeners rate recordings without seeing Akousa’s marks. The plan is written; it hasn’t run.
- How many real slips we catch. Catch rates confirmed by listeners are still being measured. They need readings marked sound by sound by ear; so far one listener has checked one sound (above).
- Native readings of this paragraph. The native ranges above are built from other sentences.
- Most first languages. Our expert check covers Mandarin speakers only, and half of them were children. For South Asian English, the scorer has now run on 2,453 Common Voice speakers (above), but no listener has marked those recordings yet; the Svarah speakers tested only the accent guess. The Tagalog and Arabic study readings are scored, but not yet marked by ear.
- Reading twice. Whether the same person gets the same marks on a second reading is not measured yet.
- Fairness by phone. Not measured yet. Women and men, and age, are measured for South Asian English speakers (above).
- Noisy rooms. Akousa won’t judge a reading when the room is too loud for it (under 15 dB signal to noise). We tested that with computer voices and recorded noise, not in real rooms.
- Where it works. iPhone first; Android is coming. American and British English only; Australian English isn’t available yet.
- Older rows. Rows marked 2 Oct were measured on 2 October 2026 with an earlier version of our lines, and are due to be measured again.
How we test
- On the phone. Akousa checks your reading on your iPhone, with a smaller copy of the speech model (226 MB) and a copy of our scorer. Before every test build, automated tests check that the phone gives the same answers as our reference scorer: every sound’s verdict in American and British English, word stress, and the result. The latest run, on 5 October, had no failures.
- The phone’s copy scores like the reference. On 100 native readings the lines never saw, the phone’s copy of the model put 96.7% of sounds in range, and the reference copy, on which the lines were set, 96.3%; 1.4% of single-sound verdicts differed. On a real iPhone, our test reading got exactly the reference result.
- Fair tests. Test speakers are picked by a fixed rule from their ID, so nobody can choose easy ones, and the lines are never tuned on them.
- Evidence first. A line moves only with numbers, recorded before and after.
- Your voice stays on your phone unless you share it. A reading is sent to us only if you tap Share under your result; shared readings help us check and improve the scoring, and a practice try is never sent. Before each test build, a check confirms that nothing else leaves the phone. The study readings above came from volunteers, through a separate page, with their consent.
Data and models we use
Thank you to the people who made these public.
- LibriSpeech: Panayotov et al., 2015. CC BY 4.0.
- CSTR VCTK Corpus 0.92: Yamagishi, Veaux and MacDonald, University of Edinburgh, 2019. CC BY 4.0.
- CMU ARCTIC: Kominek and Black, Carnegie Mellon University, 2003.
- speechocean762: Zhang et al., 2021. CC BY 4.0.
- Svarah: AI4Bharat, 2023. CC BY 4.0.
- VoxPopuli: Meta. CC0.
- Common Voice, South Asian English: Mozilla, through the Mozilla Data Collective. CC0.
- The speech model: wav2vec2 phoneme recognition, Meta AI. Apache 2.0.
- The accent recogniser, in testing: CommonAccent, Zuluaga-Gomez et al., 2023. MIT.
Questions
If a result looks wrong, or you want to check a number, write to us at hello@akousa.app. A person reads every email.