First, a question: as a learner, what kind of feedback is good feedback for me?
I’m reminded of when I learned to drive and got my driver’s license during college. I’m not afraid to ask questions—I actually want to ask them. I like to understand things clearly; only then can I feel at ease. Otherwise, feeling uncertain upsets me.
For a long time, I’ve been exposed to AI-related content. In fact, feedback is a lot like AI training: when I do something, I want someone to tell me if I’m doing it right, if my thinking is correct. But the difference—and what’s more important—is that I want to know if there’s a better way, a trick. Yes, a trick.
AI training only needs to be told right or wrong; the model will figure out how to improve on its own. What I want, though, is for someone to show me a better path, overlooked details, a simpler way of thinking—not just judging the result, but helping me open up more possibilities.
Because the model has plenty of “time” and “chances” to try randomly or iteratively until it gets better. I don’t have that ability. Perhaps we humans, as “models,” are far more complex and deep; our learning happens deep inside, even in the subconscious.
It’s like learning to ski: once I get the position of my feet and my center of gravity right, I can at least ski in control, instead of rushing down out of control. These two kinds of “skiing” are totally different, haha.
In Atobe repeater , I’ve added an evaluator system. It’s still an experimental feature, but I believe it will be a useful design, and I will definitely keep improving it.
In the past, I’ve seen many AI features that are convenient and useful. But they’re still different. What we learn is for communicating with people—now and in the future. So why not let real people evaluate directly? Is what AI says is “good” or “bad” really always good or bad?
Here’s how it works now: learners submit the audio they want to imitate, plus their own recorded imitation audio (audio only for now; more formats will be supported later), along with details like audio length, target language, skill name, and possibly the learner’s main language. All this is sent as a task.
Evaluators in the task hall can take language-related tasks at a proficient or native level. Important: tasks have a time limit. Unreviewed tasks are returned to the hall when time runs out.
So evaluators, feel free to be decisive and bold—your feedback will definitely help learners a lot.(Of course it won’t be perfect, because of the time limit. What can you do? :p)
I think sending out your own imitation is just as important as receiving feedback. Once you submit the evaluation task, you’re already halfway to success. Sometimes even just comparing carefully by yourself lets you see where to improve.
And having evaluators at a proficient or native level listen, imitate, point out issues, or give encouragement if you’re doing well is amazing. Sometimes it’s even a learning experience for the evaluators too.
I think we live in a world of the “low-precision planet concept” [1]. Many things aren’t perfectly precise or rigorous—they’re imperfect, yet they still run gently and smoothly, right?
Good enough is enough, for both learners and evaluators.
I have more thoughts on the scope of what evaluators can assess. For example, mutual help between learners of the same language should also be useful, but the format will definitely be different—it still needs more design.
Of course, as stated in my plans, I will also add AI features as a premium/advanced option. I want it to be non-essential, but still helpful for those who need it.
Finally, I have one quick question.
For a task where we judge whether someone else’s statement is right or wrong, which term do you prefer:
“Evaluation Task”, or the more comprehensive “Review Task”?
I’d really appreciate advice from native English speakers (or not native :p ).
I know tastes differ, but I’ll make the final call—and I’ll happily take on the role of the “decision-maker”! 😄
[1] Good Enough — The Principle of Tolerance for Ambiguity in Many Fields