From Dar to Nairobi: dialect coverage

People often talk about Swahili as if it were one thing. It is, in the sense that a trader in Dodoma and a student in Nairobi can understand each other. But listen for a minute and the differences are obvious: the words people choose, the rhythm of their sentences, how they say certain sounds, and which English or Sheng words slip in.

A voice model that only knows one variety will sound foreign to everyone else, or misunderstand them. This post sets out how we think about dialect coverage at Waridi, starting from Tanzania and moving outward.

Four broad varieties we plan for

Swahili varies in many more ways than four groups can capture. But for planning data collection, we start with these broad areas.

Dar es Salaam

Dar is fast, urban and mixed. Speech here is shaped by people from every region of Tanzania, plus heavy contact with English through business, media and technology. Expect quick delivery, lots of code-switching, and slang that moves fast. Bongo Flava and urban radio have spread many Dar expressions across the country.

A typical exchange:

— Mambo vipi? — Poa tu. Niko kazini, tuongee baadaye.

("How are things?" "Just fine. I'm at work, let's talk later.")

The coast and Zanzibar

Kiunguja, the Swahili of Zanzibar Town, is the historical basis of Standard Swahili. Coastal speakers, in Zanzibar, Tanga and down to Kilwa, often have a more measured rhythm and keep vocabulary and politeness forms that are less common inland. Arabic loanwords can be more frequent, and greetings tend to be fuller.

Across the border, Kimvita, the Mombasa variety, has its own distinct sound and words. It is close to coastal Tanzanian Swahili in many ways, and quite different from Nairobi.

Inland and upcountry Tanzania

Most Tanzanians speak Swahili alongside a first language such as Sukuma, Chaga, Haya, Nyakyusa or Gogo. That first language shapes pronunciation. Speakers may merge or swap certain sounds, for example r and l, or add vowels differently at word boundaries. Tone and stress patterns carry over too.

These speakers are not speaking "incorrect" Swahili. They are the majority of Tanzanian Swahili speakers, and a product that struggles with them fails most of the country.

Nairobi

Nairobi Swahili is closer to English in vocabulary, often simpler in grammar in casual speech, and full of Sheng: a fast-changing mix of Swahili, English and other Kenyan languages. A Nairobi caller might say "Niko na shida na hii simu" in standard-ish Swahili, or something much more Sheng-heavy that a Dar listener would find hard to follow.

Where the differences show up

For speech technology, dialect variation appears in a few concrete places.

• Vocabulary. Different words for the same thing. Tanzania and Kenya often borrow differently from English, and slang terms rarely travel unchanged.
• Pronunciation. Vowel quality, consonant substitutions influenced by first languages, and how loanwords are adapted.
• Prosody. Speed, rhythm and intonation. Coastal speech and Nairobi speech have noticeably different music to them.
• Code-switching patterns. How often speakers switch, and which language supplies which kinds of words. We go deeper in Code-switching is not an edge case.

For Waridi Sikia, our speech recognition model, these differences affect accuracy. For Waridi Sauti, our voices, they affect whether a voice feels local. A bank in Dar may want a voice that sounds like Dar. A broadcaster in Zanzibar may want something else entirely.

How we sample

Good coverage does not happen by accident. If we recorded whoever was easiest to reach, we would end up with mostly young, urban, university-educated speakers from Dar.

So we set targets before collecting:

1. Region quotas. Each broad area gets a planned share of speakers, not just of hours. Many hours from a few people is worse than fewer hours from many.
2. First-language balance. Within inland Tanzania, we track speakers' first languages to avoid over-representing one group.
3. Age and gender balance. Older speakers and women are often under-represented in speech data. We plan for them explicitly.
4. Rural as well as urban. We are setting up collection outside major cities, not only in Dar, Arusha and Nairobi.
5. Conversation, not only reading. Read prompts flatten dialect. We include free conversation and role-played customer calls.

All of this happens within our consent and payment rules, described in How we collect speech data with consent.

Our coverage roadmap

These are goals, not achievements. We will update them as we learn.

• First: strong coverage of Tanzanian Swahili across Dar, the coast and Zanzibar, and the main inland regions, for both recognition and synthesis.
• Next: Kenyan Swahili, including Nairobi speech and Sheng, and the Mombasa coast.
• Then: Uganda and other parts of East Africa where Swahili is a second or third language.
• Throughout: per-region evaluation, so we can see where the models are weaker, as we describe in Building a Swahili ASR benchmark.

Why start here

We are starting in Tanzania because Swahili is part of daily life there in a way it is almost nowhere else. Read Why start in Tanzania for the longer argument. But the aim from day one is a model that sounds right from Dar to Nairobi, and understands the person on the line wherever they grew up.