Insights · guide
Test voice commands with Jozi accents and noise
Drafted with Insight · edited by The WE ARE VR team
design

I trust a voice command only after it survives a noisy Joburg room.
I wrote this because phone prototypes often work beautifully beside a quiet desk. The same command can fail when people speak naturally, switch languages, or compete with coffee grinders and event chatter.
What's the problem?
Voice testing often starts with one person repeating one phrase in one room. That proves the microphone works. It does not prove your interaction works for the people who will use it.
Johannesburg speech is varied. Your audience may pronounce English words differently, mix languages, shorten phrases, or use local place names. Background noise also changes how a phone hears consonants and word endings.
The common mistake is changing the software after every failed attempt. Soon, nobody knows whether the improvement came from the command wording, the room, the speaker, or luck.
You need a small test that stays the same each time.
What's the solution?
Make a phrase list and a simple test sheet before changing your prototype. A phrase list is the exact set of commands and natural alternatives you expect people to say.
Start with five commands. Keep each one short and tied to one action. For a voice-first AR guide, you might test:
- “Show the next step.”
- “Repeat that.”
- “Take a photo.”
- “Open the map.”
- “Find Braamfontein.”
Add one or two likely alternatives beside each command. Someone may say “say that again” instead of “repeat that”. Do not add every possible sentence. That becomes a language project rather than a useful prototype [and your spreadsheet will develop its own personality].
Use the same phone, app version and phrase order in three places. Test first in a quiet office. Then try a café with normal conversation. Finally, create an event-like sound level using several nearby voices and music at the volume you expect on site.
Ask at least three people from your intended audience to participate. Do not coach their pronunciation. Let each person read the command once, then say it again in their own words.
Record four things: the phrase spoken, whether the correct action happened, how long it took, and what the system heard. Speech recognition means software turning spoken words into text or actions. If your tool shows a transcript, save it. A wrong transcript often explains a wrong action.
For example, if “Find Braamfontein” works in the office but fails in the café, do not immediately train a custom model. First try a clearer command such as “Open Braamfontein map”. Test both phrases again in all three settings. The cheaper wording change may solve the real problem.
This is the useful habit: change one thing, then repeat the same test.
What's next?
- Choose five commands from your current XR prototype and write two natural alternatives for each.
- Test them with three Jozi speakers in an office, a café and an event-like room.
- Mark repeated failures, change one phrase, then run the full test again before touching the code.
Share your test sheet, results or stubborn command with the Jozi XR community.
The WE ARE VR team
