Feedback in Voice Interactions
A comparative study of explanatory and reflective feedback when people make mistakes in voice-only learning.
Timeline
3 Weeks
Jan 2026
Focus
Voice interaction
Feedback design
Controlled experiment
Data analysis
Team
3 Designers
Guide: Prof. Swati Pal
Tools
ChatGPT voice mode
Cursor, Lovable, AI Studio
NASA-TLX
When a voice assistant tells you that you got it wrong, how should it say it? We ran a small experiment with 16 people learning personal finance by voice, comparing feedback that explains the answer with feedback that makes you think it through.
The brief & approach
01. The brief
Voice assistants are being used to teach things, but when you get something wrong in a voice-only conversation there's nothing on screen to look back at. The message is gone as soon as it's spoken. Most systems either correct you or cheer you on, and there's little work on how that choice affects how fast people recover and how much they learn. So our question was simple: how should a voice system respond when someone makes a mistake?
02. What we did
We read up on six papers about error feedback, then ran a small controlled experiment. A voice tutor built in ChatGPT asked 10 personal finance questions out loud. When someone got one wrong, it responded in one of two ways:
• Explanatory + encouraging: tells you why the right answer is right, and reassures you.
• Reflective: asks you questions so you work out the mistake yourself.
• 16 people, 8 per style, each tried only one. We measured recovery time, retries, confidence, workload (NASA-TLX) and a before-and-after quiz.
Project timeline
Phase 1
Literature review
Found the gap
Question and hypotheses
Phase 2
Experiment design
Two feedback styles
Pre and post quiz
Phase 3
Prototype
Voice tutor in ChatGPT
Prompt iterations
Final Phase
Run and analyse
16 participants
Results and write-up
Experiment design
The setup
A between-subjects experiment. 16 people, 8 per feedback style, and each person only tried one. Everyone took an 8-question finance quiz before and after, with 10 spoken questions in between. Participants were aged 22–30 and comfortable with voice interfaces.
Variables
Independent variable
Type of error feedback: explanatory + encouraging, or reflective.
Dependent variables
Pre and post test score, cognitive load (NASA-TLX), confidence, recovery time, retry count.
Control variables
The questions for the voice tasks and the AI chatbot.
What we expected
Explanatory feedback would be faster, with fewer retries and a lower workload. Reflective feedback would be slower, but help people remember more.
Approach
Outcome
Faster recovery
5.1s
Explanatory, vs 7.3s for reflective. p=0.0003, the strongest result.
Fewer retries
6.3
Explanatory, vs 11.8 for reflective. p=0.009.
Better learning
+1.5
Reflective learning gain, vs −0.1 for explanatory.
p=0.047
Not significant
47.5
Task load for explanatory, vs 55.4 for reflective. Leaned that way, p=0.142
Not significant
+0.6
The confidence gain showed not much difference. 0.5 vs 1.1
03. What we found
Three of the five measures came out statistically significant. Explanatory feedback won on speed and retries. Reflective feedback won on learning. Confidence and workload leaned towards explanatory but weren't significant, which is probably down to having only eight people per group.
04. What it means for design
There's a trade-off. Explanatory feedback is quicker and easier to deal with, so it suits short tasks and beginners. Reflective feedback makes people work harder, but they remember more. A voice tutor could start with explanations and bring in reflective questions as people get comfortable.
The full write-up has the literature review, the prompts, every chart and the complete analysis.
Read the full study on Behance →
