Dialect-aware speech recognition
Models that hear Dar es Salaam, Zanzibar, Mombasa and Nairobi Swahili as distinct, valid ways of speaking — not as noise around one standard.
Frontier models for how Africa actually speaks.
Waridi is an AI research lab training models for human–computer interaction, built around dialects, accents, code-switching and local context. We’re starting with voice, and with Swahili.
Research areas
What we’re working on.
Expressive speech synthesis
Voices with correct stress, clean vowels and natural prosody, that can carry warmth, urgency or calm the way a native speaker would.
Code-switching
“Nataka ku-check balance yangu” is a clear sentence. We model Swahili–English mixing and Sheng as everyday speech.
Real-time conversation
Streaming recognition, turn-taking and synthesis fast enough for live phone calls, with a target response time under 300 ms.
Ethically sourced speech data
Every recording we train on comes from speakers who gave informed consent and were paid fairly, region by region.
Evaluation that reflects reality
Benchmarks built from phone-quality audio, noisy streets, regional accents and code-switching — the conditions people actually speak in.
Data
Better data, collected the right way.
The most durable advantage in African speech AI is high-quality, consented data from the people who speak these languages every day. We build that dataset ourselves.
How we collect speech data with consent →- Consented
- Speakers understand what their voice is used for and agree to it in plain Swahili first.
- Paid fairly
- The rate is stated before every session, set against local wages for similar skilled work.
- Regional
- We record and evaluate region by region instead of blending one “African” accent.
- Revocable
- Speakers can withdraw at any time; withdrawn recordings are never used in new training.
Evaluation
How we’ll measure progress.
These are the targets we hold ourselves to. We’ll publish results, and the benchmark itself, when they’re ready.
AreaWhat we measureStatus
Speech recognitionWord error rate on Tanzanian phone audio, incl. code-switchingBenchmark in progress
Speech synthesisNaturalness rated by native Swahili listenersListening panels in progress
Voice agentsEnd-to-end response latencyTarget: under 300 ms
CoverageTanzanian → Kenyan → regional Swahili, then more languagesTanzanian Swahili first
Writing
All articlesResearch notes.
- 8 September 2026Designing voices that sound TanzanianHow we approach casting, recording and shaping Waridi Sauti voices so they sound like someone from Tanzania, not a generic accent stitched together from everywhere.Research
- 25 August 2026Code-switching is not an edge caseTanzanians mix Swahili and English mid-sentence every day. Here is why Waridi Sikia and Waridi Sauti treat that mixing as normal speech, and how we build for it.Research
- 11 August 2026How we collect speech data with consentEvery recording we train on comes from a speaker who understood the request, agreed to it in their own language, and was paid fairly and on time.Ethics
- 28 July 2026Latency budgets for real-time voice agentsA voice agent has roughly a breath's worth of time to answer. Here is how we split that time across listening, thinking, speaking and the phone network.Engineering
- 14 July 2026What good Swahili speech sounds likeStress, clean vowels, noun-class agreement and the way Tanzanians read numbers, prices and times. The details that decide whether Swahili text-to-speech sounds natural.Research
- 9 June 2026Building a Swahili ASR benchmarkBefore we can claim Waridi Sikia works, we need a test that reflects real Swahili speech: phone lines, noisy streets, regional accents and constant code-switching.Research
- 19 May 2026Voice cloning and consentA cloned voice can open doors for creators and businesses, or be used to deceive. Here is how we are designing Waridi's cloning so only the voice owner decides.Ethics
- 17 March 2026From Dar to Nairobi: dialect coverageSwahili sounds different in Dar es Salaam, Zanzibar, Mombasa, Mwanza and Nairobi. Here is how we plan to cover that variation instead of flattening it.Research
- 10 February 2026A developer tour of the Waridi APISpeech synthesis, recognition, streaming, agents, cloning and dubbing, all from one API. A hands-on walk through the endpoints with curl, Python and Node.Engineering