YZ Yao Zhang
← All research

PROJECT 03 / AI reliability

East Asian Users’ Experiences in Voice Interaction with Generative AI

When voice AI misunderstands a request, what do users change, and how do those experiences shape their trust?

Role
Independent researcher
Project period
Aug 2026–In progress

01 / STUDY DESIGN

Who took part

I recruited 24 East Asian speakers who use English as a second language. The protocol used the same set of 10 scripted ChatGPT Voice tasks for each participant, split into two phases.

08Mandarin native speakers
08Korean native speakers
08Japanese native speakers

02 / STUDY PROCEDURE

From first request to replay

After the scripted opening request, participants chose whether and how to repair the interaction. Each task had a two-minute limit, and participants could stop earlier.

  1. SETUP

    Consent and setup

    Check recording, microphone, and screen sharing. Try one practice interaction.

  2. T0

    Baseline trust

    Rate confidence from 1 to 10 and discuss prior voice AI use.

  3. PHASE 1

    Tasks 1–5

    Start each task with a scripted English request in a fresh conversation.

  4. T1

    First trust check

    Rate confidence again and explain the rating before a break.

  5. PHASE 2

    Tasks 6–10

    Continue with five new tasks under the same procedure.

  6. T2

    Final trust check

    Repeat the confidence rating and ask which experiences shaped it.

  7. REPLAY

    Stimulated recall

    Replay recorded interactions in a semi-structured interview about what participants noticed and why they acted.

03 / RESEARCH METHODS

How I analyze the evidence

01 / INTERACTION RECORDING

Behavior coding

Compare actual speech with the script, visible platform text, and AI response. Code observable responses and repairs separately.

02 / STRUCTURED TRUST CHECKS

Descriptive statistics

Summarize response patterns, repairs, and task outcomes. Examine 1–10 confidence ratings and their explanations at T0, T1, and T2.

03 / PLAYBACK INTERVIEWS

Reflexive thematic analysis

Analyze why participants repeated, rephrased, switched to typing, or stopped, and which adjustments they described as reluctant compromises.

I compare observed repairs with participants’ own accounts. Completing a task, choosing to rely on the system, and reporting confidence are examined separately. The project is ongoing.