I wanted my children to talk to AI about the textbooks they actually use at school without needing a phone. A seemingly simple voice interface became a product problem spanning knowledge, hardware and everyday experience.
From family feedback to a concrete change
English too fast; calling another assistant did not switch
Role-specific speed; switching after interruption
Repeat the interaction and inspect the result
Based on prototype iteration records. This shows how feedback changed the design, not that every noise or distance condition has been solved.
01 / Conversation alone is not enough
The first version was slow, sounded poor and was sensitive to background noise. Receiving a new instruction while the device was speaking was difficult, leaving family members waiting for the system.
A speakerphone with full-duplex audio and echo cancellation improved the audio conditions. But simultaneous playback and capture do not automatically deliver software interruption. Wake detection, cancellation and role switching still need their own handling.
02 / Device resources and voice quality
The Raspberry Pi’s resources constrained local speech models. To get a more natural voice, I introduced cloud speech services while retaining fallback paths. Network access, service availability and cost became part of the product.
Running on a local device is different from doing all processing locally. The experience depends on the whole chain; a speech service’s first audio chunk is not a measure of end-to-end conversation latency.
Voice interaction and background work
Capture, playback and wake detection
Assistant selection and active dialogue
Sources, models and speech services
MyAgent handles selected background tasks
03 / Let actual use change the design
“The English is too fast” led to role-specific speed adjustments. Calling another assistant during speech exposed a switching problem after interruption. These small observations gave a clearer next step than a longer feature list.
Textbook integration also needs visible evidence. The demo puts responses alongside source material so viewers can inspect how it is used; that is not a measured educational outcome.
From prototype to an easier product
This remains a home prototype without public installers or source code. The next direction is simpler setup and validation on devices beyond Raspberry Pi. Audio, permissions, dependencies and updates need to be addressed on each operating system.