Names and identifying details of every student here are changed, as they are everywhere on the blog.
Arina is twenty-eight, lives in Albany, and had a green streak of ninety-one days when she wrote to me. My pronunciation is rated excellent, she said, and I cannot order coffee. Both halves were true.The tool scores you highest when you are doing the least useful thing
The apps weren’t lying to her. They were measuring something real, and measuring it carefully, and the number they returned was honest. It just was not a number about the thing she wanted.
Reading a sentence you can see is the easiest thing you will ever do in Russian, and it is the thing your app is best at grading.
Those figures come from English, not Russian, and I am not going to pretend otherwise. But the shape of the problem transfers, and Russian gives it more to work with: unpredictable stress, consonant clusters, and reduction that swallows vowels the moment you speak at normal speed.What each kind of tool is actually for
Our review of the apps themselves covers Duolingo, Babbel, Pimsleur and the rest in detail, so I won’t repeat it here. This is about the newer layer sitting on top of them.| Tool | Genuinely good at | Quietly bad at |
|---|---|---|
| Chatbot, text | Unlimited written practice, grammar on demand | Remembering what you keep getting wrong |
| Chatbot, voice | Lowering the fear of speaking at all | Judging whether you were understood |
| Pronunciation scoring | Single sounds and read-aloud lines | Anything you compose yourself |
| Dictation tools | Showing what a machine heard | Telling you why it heard that |
Nobody uses the fourth row, and it is the one that works.
The fourth row is the useful one and almost nobody treats it that way. A dictation tool is not a grader. It is a very literal listener, and reading its transcript of your own speech is closer to real feedback than any score out of a hundred. That is not what I said is the most useful sentence a learner can produce in front of one. One doctor I taught for a year used the voice mode of a chatbot every evening on his commute and told me in month eight that he had stopped, because he could not tell whether he was improving. Nothing in the tool had told him he was wrong about anything. He had been having pleasant conversations with something that agreed with him at speed, and the agreement was the problem. It never once asked me to repeat myself, he said, and that was the sentence that made him quit.A tool that never makes you repeat yourself is not listening. It is guessing.
Two prompts that turn a chatbot into something sharper
What Arina did instead
She kept the apps, because the streak was doing no harm and she liked it. I asked her to add one thing on top, which took about four minutes a day and required nothing she did not already own.Week two of the new routine, group lesson
Arina I recorded myself ordering coffee and the transcript said кофе с молоком, but I said с молоко.
Me So what did the machine hear that I would also have heard?
Arina …that I dropped the ending. Every time. Even when I know it.
Nobody corrected her. She read a transcript of her own voice and found it herself, which is why it stuck.
A pronunciation score tells you how you did. A transcript tells you what you said.
Arina did this for six weeks and brought me a page of differences. Most of them were endings, and most of the endings were the same three. She had known all three rules since her first year, and had been dropping them in speech for two years without a single app ever mentioning it, because every sentence she read aloud to a scoring tool had the endings printed in front of her.Where this stops working
Being understood and being correct are different things, and only one of them can be measured.
Software stops helping at the point where those two come apart, and in Russian they come apart constantly. Your transcript will tell you the machine heard в магазине when you meant в магазин. It won’t tell you that a person would have understood you anyway, or that the sentence you produced is grammatical and something no Russian speaker would say. Our earlier piece on the apps and their feature lists makes the same point from a different direction. For that you need somebody who has heard a lot of Russian and can say nobody would put it that way without being able to cite a rule. There is no version of this where the software gets there on its own. At Palme School we teach online, in groups of up to eight, two forty-minute classes a week, and the first two lessons are free. You can book those here, and there are printable materials for learners in our resource library, plus a conversation school for people whose reading has run ahead of their speaking.Frequently asked questions
They are most accurate on read-aloud lines and least accurate on real speech, which is the opposite of what a learner needs. A high score on a sentence you can see says very little about a sentence you have to build.
For written practice and grammar, a general chatbot with a well-built prompt. For speaking, a plain dictation tool used as a mirror rather than a grader. For the apps themselves, our separate review compares them by task and budget.
Partly. Recording yourself and reading the transcript gets you a real error list, and that is more than most learners have. What it can’t do is tell you when a correct sentence still sounds wrong.
Because the app measures you under the easiest possible conditions. Prepared text, no time pressure, no other person waiting. Remove those three and the task changes entirely.
Sources
- Cornelissen, S., et al. (2023). Automatic speech recognition of non-native child speech for language learning applications. arXiv preprint. arxiv.org
Palme School: Russian for bilingual children · Conversation school · Free materials
Lily Rozberg has been teaching for sixteen years, first English in Russia and now Russian to bilingual children across the US and Canada. She writes for Palme School about heritage language, reading, and the things report cards do not measure.
