Voice capture and
speech recognition
Unity Microphone API; Baidu
Short Speech Recognition
Professional API; Mandarin
near-field fast recognition
model (https://vop.baidu.com/pro_api; dev_pid 80001)
{
“audio”: “16-kHz mono WAV (Base64)”,
“auth”: “access token”,
“device_id”: “local device ID”
}
{
“api_output”: “transcript text”,
“local_file”: “recorded_audio.wav (overwritten)”,
“next”: “dialogue module”
}
Teachers supervised
recording with researcher
support
Recommended: a clear
microphone-use notice and deletion of temporary audio
after transcription, unless
retention has been approved
Guided child and
AI dialogue
DeepSeek-V3.2-Exp (non-
thinking mode) API; model
identifier deepseek-chat; POST
https://api.deepseek.com/v1/chat/completions
{
“system_prompt”: “child-friendly drawing interview”,
“messages”: [
“current transcript”,
“dialogue history”
],
“image”: null
}
{
“api_output”: “assistant text”,
“local_files”: [
“chatMessage.txt”,
“prompt.txt”
],
“next”: “speech synthesis”
}
A child-friendly drawing-
interview prompt guided the
dialogue. Teachers
supervised the sessions
with researcher support
Recommended: personal-data
redaction, input/output
screening, explicit limits on
dialogue turns, and safe
fallback responses
Conversation summarization
and animation-prompt creation
DeepSeek-V3.2-Exp (non-
thinking mode) API using the
same deepseek-chat identifier
and endpoint as the dialogue
module
{
“dialogue”: “text from prompt.txt”,
“instruction”: “create a concise image-to-video prompt”
}
{
“api_output”: “motion prompt”,
“local_handling”: [
“application memory”,
“diagnostic output”,
“Python argument”
]
}
The dialogue was summarised
into an animation prompt.
Researchers reviewed the
generated animations before
children viewed them
Recommended: removal of
identifying details and explicit
constraints limiting animations
to simple movements
Speech synthesis
Baidu Short Text Online Synthesis API.
Local Flask endpoint https://localhost:5000/generate_voice;
Baidu OAuth endpoint
https://aip.baidubce.com/oauth/2.0/token;
Baidu Text-to-Speech endpoint
https://tsn.baidu.com/text2audio.
The configured output was Chinese speech in
MP3 format (aue = 3), using the base
voice Du Yaya (度丫丫; per = 4)
{
“text”: “assistant response”,
“parameters”: [
“voice”,
“speed”,
“pitch”,
“volume”
],
“child_audio”: null,
“drawing”: null
}
{
“api_output”: “MP3 stream”,
“local_handling”: “immediate Unity playback”,
“saved_as_research_record”: false
}
/
Recommended: localhost-only
binding, disabled debug mode,
text-length validation, timeouts,
and data-retention controls
Drawing-to-animation
generation
Alibaba Cloud Model Studio
DashScope VideoSynthesis
through the DashScope Python
SDK; model Wan 2.2 Image-to-
Video Plus (wan2.2-i2v-plus);
output resolution 1080P
{
“image”: “child drawing (Base64 data URL)”,
“prompt”: “motion prompt”,
“model”: “wan2.2-i2v-plus”,
“resolution”: “1080P”
}
{
“api_output”: “temporary video URL”,
“local_files”: [
“video_result.txt”,
“video.MP4”
],
“use”: “local Unity playback”
}
Researchers reviewed
generated animations
before they were shown to
children. The images
submitted for animation
generation were
children’s drawings
Recommended: checks for
faces and identifying details,
additional prompt/output
screening, and disabling the
logging of temporary URLs
Local files, logs, and
research records
Local Windows filesystem
and Unity diagnostic logs
{
“external_transmission”: null,
“local_files”: [
“recorded_audio.wav”,
“chatMessage.txt”,
“prompt.txt”,
“video_result.txt”,
“video.MP4”,
“logs”
]
}
{
“uses”: [
“transcription”,
“dialogue continuity”,
“prompt creation”,
“video playback”
],
“public_release”: false
}
Files were stored on the
local filesystem for
transcription, dialogue
continuity, prompt creation,
and playback
Recommended: pseudonymous
identifiers, access-controlled
encrypted storage, and a
documented data retention
and deletion schedule
Consent, disclosure,
and human oversight
No external API. This row
describes the supervised
research procedure
reported in the manuscript
{
“external_transmission”: null,
“research_inputs”: [
“guardian written consent”,
“child assent”
]
}
{
“system_role”: [
“expressive scaffold”,
“no grading”,
“no diagnosis”,
“no teacher replacement”
]
}
Written consent was obtained
from guardians. Children’s
willingness to participate was
confirmed before each activity,
and withdrawal was permitted.
Botune’s AI role was disclosed,
and teachers and researchers
supervised the sessions
/
Credentials and network
security
Authentication credentials
for DeepSeek, Baidu, and
DashScope; HTTPS provider
endpoints; and a local Flask
endpoint restricted to localhost
after the study
{
“authentication”: “provider-specific access token transmitted only to the corresponding HTTPS endpoint”,
“credentials_in_content_payload”: false
}
{
“credential_storage”: “access-controlled runtime configuration outside the source repository”,
“authentication_logging”: false,
“temporary_url_logging”: false
}(Post-study configuration)
Requests to external serviceproviders
used HTTPS
Credentials were revoked and
rotated, removed from source-
code files, and moved to
access-controlled runtime
configuration outside the
source repository.
Authentication information was
excluded from logs; temporary
media URLs were not publicly
released; Flask was restricted
to localhost