Translation experiences for the multilingual room

If you ask someone on a street in Nairobi what language they speak, they won’t name just one. The answer is almost always three.

They have the language they grew up speaking at home — like Luo, Kipsigis, or Kamba. Then there’s Swahili, which gets them through the market, a conversation with a stranger, or a commute on a matatu. Finally, there’s English for the office, spreadsheets, and official paperwork. Because schools teach both Swahili and English, most people have a functional grasp of both. But the home language is what they absorbed before school ever started. It’s the one tied closest to their identity.

This multi-layered setup isn’t an anomaly; it’s the global default. What makes it fascinating from an engineering perspective is that these three tongues aren’t even distant cousins. Luo is Nilotic. Swahili is Bantu. English is Germanic. A typical person walking through Nairobi routinely jumps across three entirely separate language families every day without viewing it as anything remarkable.

Mainstream communication software completely misreads this environment. Silicon Valley design principles rely on a monolingual bias: people meet, pick a single language, and stay there.

But real multilingual rooms don’t operate on a single track. People code-switch mid-sentence, selecting words based on shifting dynamics — formality, intimacy, or who just walked through the door. It is entirely common to watch two people hold a fluid conversation where Person A speaks only Swahili, Person B responds entirely in Luo, and neither feels the need to switch.

We wanted to see what translation looks like when it’s built for natural, multi-directional rooms instead of rigid one-way bridges. We put together some applications using a deliberate mix of high- and low-resource languages reflective of Kenya’s actual landscape: English, Swahili, Kikuyu, Kamba, Kimeru, Luo, and Somali.

Scenario 1: Two people, one phone

Take the simplest setup: two people sharing one phone. You select your language, choose what the other person needs to hear, and tap the microphone. The phone speaks the translation aloud, and you hand it over for the reply.

[embed]

The layout isn’t novel, but the language pairs are. Standard commercial tools handle English-to-Swahili fine, but our prototype bridges Swahili directly to Kikuyu. Most existing platforms either ignore this pairing or route the data through an English intermediary, which strips out regional nuance. When two people share an imperfect second language, a direct local-to-local bridge fundamentally alters the dynamic.

Scenario 2: A room with several languages

The mental model shifts entirely when you move to a multi-device room. One user creates a session using a QR code, and everyone else joins on their own phone, selecting their preferred language for reading and listening. There is no designated “meeting language.”

When someone speaks, the system transcribes their words in their native tongue while broadcasting real-time translations to every other screen. Each phone displays a personalized, canonical view of the discussion. The software assumes the burden of the bridge, leaving the participants free to speak naturally.

This changes the math for everyday situations. At a family gathering spanning three generations and four distinct comfort zones, the youngest teenager doesn’t have to spend the afternoon acting as a live translator; they can just participate. A community health worker fluent in Swahili can talk with a household that only speaks Kikuyu at the speed of a normal conversation, without waiting for clunky, sentence-by-sentence human interpretation. Local field teams and international NGO staff can stop performing the awkward ritual of pretending everyone is equally comfortable in English.

By making every participant a simultaneous source and target, the technology adapts to the people in the room rather than forcing them to bend to the limits of a tool.

Scenario 3: One speaker, many readers

We see a similar shift in one-to-many broadcasts like town halls, workshops, or sermons. The speaker’s live words anchor the top of a main screen or individual devices, while a real-time grid of multiple language translations streams below it.

[embed]

This is where localized language models prove their worth. Think of a pastor preaching in Swahili to a congregation split between Luo, Kikuyu, and Somali speakers, or a teacher running an orientation night that needs to land clearly across three different linguistic groups. Human interpretation is wonderful, but it’s too expensive for routine events, so most communities simply go without it. Creating a lightweight, automated accessibility layer turns a fleeting spoken presentation into a permanent, multilingual record that attendees can save and review later.

None of these environments ask a user to abandon their language for someone else’s. There is no hierarchy, and no designated “primary” tongue. The design challenge here isn’t about adding more features; it’s about making the technology disappear into the natural cadence of human speech. When the interface works, it functions as an accessibility layer for natural multilingualism, and language differences stop acting as a gatekeeper for who gets to participate.

Polyglot and Podium are research prototypes from Thiomi. They are available at drlugha.com/translate.