Designing voices that sound Tanzanian
How we approach casting, recording and shaping Waridi Sauti voices so they sound like someone from Tanzania, not a generic accent stitched together from everywhere.
A synthetic voice is a set of choices. Who you record, what you ask them to read, how you ask them to read it, and who decides whether the result is good. Every one of those choices shows up in the final audio. This post explains how we are making them for Waridi Sauti, our text-to-speech family.
Our goal is simple to say and hard to do: when someone in Dar es Salaam, Mwanza or Mbeya hears a Waridi voice, it should sound like a person from here. Not a foreigner reading Swahili carefully. Not a vague "African" accent. A Tanzanian speaker.
Start with people, and with consent
Every Waridi voice begins with a real voice actor who has agreed, in writing, to have their voice modelled. We explain in plain Swahili what the recordings will be used for, how long the voice will be available, and how they can withdraw. We pay for the recording sessions, and we design our agreements so that speakers share in the ongoing use of their voice rather than signing it away once.
We cover the data side in more detail in how we collect speech data with consent. The short version: if we can't explain the deal to a speaker in a few sentences, the deal is wrong.
Casting for a place, not a continent
There is a temptation in speech tech to build one "African English" or one "Swahili" voice and call the job done. That produces something nobody actually sounds like.
Swahili in Tanzania carries regional colour. A speaker raised in Zanzibar or Tanga has a different melody from someone raised in Arusha or Dodoma. Many Tanzanians grew up speaking another language at home, such as Kisukuma, Kichaga or Kihaya, and that shapes their Swahili in small, recognisable ways. None of these is wrong. But a voice should be one of them, consistently, not a blend of all of them.
So when we cast, we write down where the voice is from and keep that identity stable across the dataset. We would rather offer several clearly located voices than one that belongs nowhere.
Persona and register
A voice used for a bank's customer line has a different job from a voice narrating a children's story. We design personas around the job, then cast for them.
The biggest axis in Tanzanian Swahili is register. Formal Swahili, the kind heard on the news or in an official letter, is careful and polished. Everyday customer service is warmer and quicker. Compare:
Tafadhali subiri kidogo wakati tunashughulikia ombi lako.
Pole sana kwa usumbufu, tunalishughulikia sasa hivi.
Both are correct. Both are polite. They feel very different on a phone call. A good agent voice needs to move between them depending on context, the way a person at a service desk does.
For each persona we define:
- Who they are: age range, region, the kind of setting they would work in.
- Default register: formal, neutral or warm.
- Pace: unhurried for instructions, quicker for small talk.
- Greetings and closings: how this person would naturally open and end a conversation, from Habari za asubuhi to Asante, karibu tena.
Recording scripts that sound like life
Most of what people say to a voice agent is not literary. It is account numbers, place names, short questions and half sentences. If the recording script is only formal prose, the voice will read everything like a news bulletin.
Our scripts mix several kinds of material:
- Balanced reading text to cover the sounds and syllable combinations of Swahili.
- Conversational turns written by native speakers, including greetings, apologies, confirmations and polite refusals.
- Numbers, dates, prices and phone numbers read in natural Tanzanian style.
- Common English loanwords and brand names as Tanzanians actually say them.
- Short expressive passages: reassurance, good news, gentle urgency.
We also leave room for the actor. When a line feels unnatural in their mouth, we change the line, not the actor. Native speakers on the session catch phrasing that reads fine on paper but nobody would say.
Expressive styles, used carefully
Expressiveness is what separates a usable voice from a pleasant one. We record controlled styles such as calm, cheerful, apologetic and serious, so that developers can choose a tone without the voice drifting into a different person.
We are cautious here. Over-acted warmth on a support call sounds fake fast. Our target is a voice that a caller would describe as mtu wa kawaida: an ordinary, friendly person.
How we decide a voice is good
Automated metrics help us catch regressions, but they cannot tell us whether a voice sounds Tanzanian. Only listeners can.
Our evaluation plan is built around native-speaker panels in Tanzania, drawn from different regions and age groups. We ask them practical questions:
- Where do you think this speaker is from?
- Would you trust this voice on a call from your bank or mobile money provider?
- Did any word sound foreign or wrong?
- Does the voice sound like it understands what it is saying?
The second question matters most to us. A voice can be technically clean and still feel distant. We want listeners to hear someone familiar.
What comes next
We are building the first Sauti voices for Tanzanian Swahili, starting with sauti-standard for everyday product use and sauti-pro for richer, more expressive work. We will add more regional voices over time, and then extend the same process to Kenya and Uganda, casting locally each time.
If you want to try the early voices, the text-to-speech docs explain how, and get early access to join the sandbox. We would love to hear what sounds right to you, and especially what doesn't.