A student records a passage on Monday and hears what was wrong with it on Friday. By then they have practised the mistake four more times.
EdTech AI automation means a submission gets a response while the student is still thinking about it. Audio is transcribed and compared against what the exercise asked for, and a score comes back in seconds. A teacher sets the standard and reads the borderline ones.
A learning platform can enrol a thousand students this week. It cannot hire a hundred teachers in the same week, and marking is the one part that does not get cheaper as you grow. Feedback slows down exactly when more people are waiting.
A correction lands after the habit has set, which is the one moment it cannot do its job.
Exercises that need marking by ear are the first ones a busy course stops assigning.
A course forum is a small community with nobody paid to read it after the first month.
A recording is transcribed and compared against the reference text, then scored on the things the exercise was set for. Your rubric decides, and a teacher can override any score.
The score comes with a sentence about what to fix rather than a number on its own. Students get it in seconds, and the teacher sees the pattern across a cohort instead of one script at a time.
A cohort forum gets read the way a moderator would read it, so an unanswered question or an unkind reply surfaces the same day. Moderation inside a course, not across a public feed.
Four steps. A teacher sits at the last one, and the standard everything is judged against is theirs.
A student records a passage or writes an answer.
Audio becomes text on hardware you choose.
The model marks it against your rubric.
A teacher checks the borderline ones.
Human checkpointOnce a platform can hear a submission and score it, several other jobs turn out to be the same job wearing a different name.
Three answers an education team needs before anything else. Student work is somebody else's child's voice, which changes what a default setting is worth.
Student audio and scores never train a model. An open weight model on hardware you own means the recordings never leave your network at all.
What gets collected from a student and how long it stays is documented before the build starts. That document is what a school asks for, and it is also what a store review asks for.
Microphone permission is requested at the moment it is needed, with the reason on screen. Where the learner is a minor, GDPR puts that consent with a parent, and the flow has to be built for it.
These are the pieces a learning build tends to need. Your own LMS is the starting point, and nothing on this list is a requirement.
Worse than on an adult reading a script, and that gap is the whole design problem. Children's speech and a strong accent are both under-represented in a general model, so a score that looks confident can be wrong. Recordings from your own learners are what close it, and the gap should be measured before anybody relies on a number.
No. It produces a score and a reason, and a teacher owns the grade that goes on a record. Anything near a pass mark is held for a person by default, because that is where a wrong call actually costs somebody something.
No. The scoring and transcription sit beside your LMS as services it calls, so the platform your students log into stays the one they know. An LMS that exposes no API costs more, and week one is where that gets established.
The data questions move to the front. What a student records and who can reach it get settled before a screen is designed. A build that guesses gets taken apart later, and a store review and a school both examine it.
Some will try, and the ones who do teach you where the rubric is loose. A model that scores a transcript can be fed a transcript, so anything carrying a grade needs the recording kept and spot checked. That is a policy decision more than a technical one.
More than most platforms expect, and the number matters less than whether it was marked consistently. A set of submissions your own teachers have already graded is what the model gets calibrated against. Where nothing has been graded yet, the first weeks build that set rather than tune against it.
Data entry, answering the same tickets, chasing numbers between systems. We automate the parts that repeat. Your team keeps the parts that need judgement.
