The most common question we hear about the Singing Carrots AI Vocal Coach is some version of this: “Is it actually adapting to me, or is it just a script with a chat window?”
We wanted to answer that with data, not a product claim. So we looked at 7 months of usage – 2,249 singers, 13,277 practice sessions, roughly 349,000 sung exercises. Here is what we found.
In our data, 91.5% of the notes the AI coach asks a singer to sing fall inside that singer’s own demonstrated comfortable range. After a successful attempt, the AI coach makes the next exercise harder 31.5% of the time; after a struggle, only 3% of the time. That 10x swing is the clearest single number for what “adaptive” means in practice.
The rest of this article explains what those numbers mean and how we got them.
What we found – at a glance
- The coach places 91.5% of its exercises inside each singer’s own comfortable range
- It reacts to every attempt: 10x more likely to increase difficulty after success than after a struggle
- The first exercise already fits your likely voice, before it has heard you sing
- A typical session introduces about 7 exercises you have never sung before, while revisiting known material in new keys
- Difficulty does not steadily increase – and that turned out to be the most interesting finding of all
Does the coach center on YOUR voice, or a generic average?
Yours. Among our 2,249 singers, 77% of all variation in where the coach places exercises is between different singers – not between a single singer’s own sessions. Who you are determines where the coach works. That is the opposite of a fixed script.
In numbers: exercise placement correlates r = 0.876 with each singer’s demonstrated comfortable range. To check whether this is just the coach “agreeing with itself” – placing exercises where it already placed exercises – we rebuilt each singer’s comfortable range using only their earlier history. The result: r = 0.801, with 90.2% of notes falling in-band. The result holds.
One detail worth noting: the correlation with a singer’s range-test midpoint is only r = 0.552. The coach targets where you sing comfortably – not the extremes you can strain to on a one-off test. Exercise centers sit slightly below the test midpoint. A teacher wouldn’t park you at your screechy top note every session. The coach doesn’t either.
How quickly does it lock on to your voice?
Faster than we expected.
The very first exercise a new singer receives already correlates r = 0.641 with their eventual comfortable range – built from their range test and profile before they have sung a single coached note. By session 3, that correlation is around r = 0.88, and the spread of exercise placements across singers has widened by a factor of 1.32, converging on the natural spread of the voices themselves.
The honest framing: it starts from an informed guess and converges onto your voice within a session or two.
Does it react to what you just did – or run its plan regardless?
It reacts.
We looked at 239,720 consecutive exercise transitions – pairs of back-to-back exercises from the same singer. Among 947 singers who had both successful and struggling moments in their sessions (65% of transition users), the pattern is consistent across the whole dataset.
After a successful attempt: the next exercise is harder 31.5% of the time.
After a struggle: harder only 3.0% of the time.
The coach also nudges the same exercise up a key after success (5.8% of the time, vs 1.4% after a struggle).
One robustness check worth one sentence: when we restrict to exercises the coach itself chose – excluding user-initiated retries – the simplify-after-failure effect nearly doubles. The reaction is the coach’s, not the singer’s.
Does it give you the same exercises over and over?
No – and it doesn’t go to the other extreme of constant novelty either.
The median session contains 13 distinct exercise patterns. About 7 of them are brand-new to that singer – patterns they have never sung before. The rest revisit known material, and 40% of those repeated patterns come back transposed to a new key. Same drill, new key is a classic teaching move: it builds the skill without the boredom of literal repetition.
Novelty does not fade with practice. Through session 10, 96-99% of sessions still introduce at least one pattern that singer has never sung before. At the same time, around 95% of sessions also revisit something known. The mix holds.
One thing we checked: the variety is not a heavy-user artifact. Among each singer’s first 20 distinct exercises, the median is 13 distinct patterns – and that number is flat across light and heavy users (correlation with total exercise volume is approximately 0).
What surprised us: the coach doesn’t just crank up difficulty
We went into the analysis expecting a clear upward ramp – exercises getting progressively harder from session 1 to session 10. That is not what we found.
Session 1 is actually the hardest – faster and wider-ranged than later sessions. It looks like an assessment: the coach starts by testing what you can do. From session 2 onward, overall difficulty holds roughly steady, while phrases get slightly longer (+0.25 notes per exercise) at a slightly calmer pace (-1 bpm).
The individual trajectories are where it gets more interesting. Among singers who kept practicing long enough to reach session 10 – about 15% of users – roughly 25% ended up on harder material than session 1, 25% on roughly the same level, and 50% on easier material.
That 50% figure is worth sitting with. It does not mean those singers got worse. It means the coach settled them onto material that matched their voice – which, for many singers, is less demanding than what session 1 probed. That is what meeting someone where they are looks like in the data.
How does this connect to our earlier results?
The effectiveness article we published earlier – the 7-months outcomes article – has data from 2,073 singers over the same period. It showed that AI-coached users improved pitch accuracy by an average of +5.9 percentage points over 4 weeks (paired analysis of 358 singers), with an average vocal range expansion of +2.8 semitones (359 singers). This article does not re-examine those outcomes. It examines the mechanism underneath them.
The adaptivity data and the outcomes data are consistent: a coach that places exercises inside each singer’s comfortable range, reacts to every attempt, and keeps material fresh is doing something recognizably different from a fixed program. Try the AI singing coach and see what it does with your voice specifically.
How we measured this
Data window: approximately 7 months (late November 2025 to early July 2026). 2,249 singers, 13,277 sessions, roughly 349,000 sung exercises. A duplicate-logging bug in the raw data was identified and removed before analysis – we mention this because it is part of how you should read the numbers.
Every statistic is computed per user first, then averaged across users – so a singer who practiced 200 times counts the same as one who practiced 5 times. All headline numbers carry 95% confidence intervals computed by resampling users.
Cohort-restricted findings name their cohort: “singers who reached session 10” is 15% of users; “singers with both good and bad moments in the transition data” is 65% of transition users. We do not report cohort findings as if they applied to everyone.
“Exercise pattern” means a derived fingerprint of the note and rhythm pattern – our definition, not a universal standard. “Difficulty” is a content-complexity index we built, weighted equally across note range, tempo, and phrase length. It measures how demanding the material is on paper, not how hard it felt.
Comfortable range describes where a singer actually sings well across their sessions – not the extremes they reached in a one-off range test. Range numbers reflect the current coach version.
These findings are observational. “Consistent with” is the right phrase; “proves” is not. What we did not find – a steady difficulty ramp – is stated plainly above, because we think it is as informative as what we did find.
FAQ
Adapting. The single clearest number: after a successful attempt, the coach makes the next exercise harder 31.5% of the time; after a struggle, only 3% of the time. That 10x swing means the coach is reacting to what you just did, not running a fixed plan. We verified this across 239,720 back-to-back exercise transitions from 947 singers.
It starts from an informed guess and converges on your voice within a session or two. Your first exercise already correlates r = 0.641 with your eventual comfortable range — built from your range test and profile before you have sung a single coached note. By session 3, that correlation reaches r = 0.88. It doesn’t need to hear you sing to make a reasonable first choice; it uses what you told it and refines from there.
Yes. Where the coach places exercises correlates r = 0.876 with each individual singer’s demonstrated comfortable range — meaning it consistently returns to where that specific singer sings well. 77% of the variation in exercise placement is between singers, not within one singer’s own sessions: who you are, and what you’ve done before, determines what you get today.
No. The median session contains 13 distinct exercise patterns, of which about 7 are brand-new to that singer. The rest revisit known material, and 40% of those repetitions come back in a different key. Novelty holds through at least session 10 — 96–99% of sessions still introduce at least one pattern you haven’t sung before. It’s not endless repetition, and it’s not constant novelty either; it’s the mix a teacher would use.
Read more
To get an overview of the AI Vocal Coach app field, read our comparison article: Top 7 AI Vocal Coaches, based on real data. For the comprehensive answer to whether AI vocal coaching works, including independent research: Do AI vocal coaches actually work? Also read: Can AI replace a vocal coach? An honest answer.
Also read: How to practice with an AI vocal coach: the session ritual