Sauti TTS: Giving Swahili a Voice in Modern AI

Two hundred million people speak Swahili.

It is the lingua franca of East Africa, the language of commerce in Dar es Salaam, of parliament in Nairobi, of daily life from Mombasa to Bukavu.

Voice is the most natural way to interact with technology, yet Swahili has been largely absent from the neural speech synthesis revolution that transformed English, Mandarin, and other high-resource languages over the past decade.

That is why we built Sauti TTS.

Sauti means voice in Swahili.

And that is exactly what this project is about: giving Swahili a natural voice in modern AI systems.

Today, we are releasing Sauti TTS, MsingiAI’s neural Swahili text-to-speech system designed for natural pronunciation, correct Swahili rhythm, realistic prosody, and real-world usability.

🤗 Model Weights: https://huggingface.co/msingiai/sauti-tts
Github Repository: https://github.com/Msingi-AI/sauti-tts

Why Swahili TTS Matters

Speech technology becomes most useful exactly where typing becomes hardest.

That includes:

• education

• accessibility

• conversational AI

• audiobooks

• healthcare systems

• customer support

• voice interfaces for mobile-first populations

But building high-quality Swahili TTS is harder than simply training a multilingual model and hoping it generalizes.

Swahili has:

• a five-vowel system

• penultimate stress patterns

• prenasalized consonants

• rhythm and timing patterns distinct from English

Most multilingual TTS systems reproduce Swahili with foreign articulation patterns.

The vowels are slightly wrong.

The rhythm feels unnatural.

The stress lands in the wrong place.

The speech sounds intelligible, but not authentically Swahili.

We wanted to build something better.

What Sauti TTS Is

Sauti TTS is a Swahili neural speech synthesis system built on top of F5-TTS v1 Base, a state-of-the-art flow matching architecture for speech generation.

At a high level, the model learns to progressively transform random noise into natural speech, refining the waveform step by step until it becomes human-like audio.

The intuition is surprisingly simple:

The model starts with noise and gradually shapes it into speech the way a sculptor refines rough stone into a finished figure.

At each stage, the system learns:

• pitch

• rhythm

• timing

• pronunciation

• breath patterns

• phoneme transitions

Those small details are what make speech sound alive rather than mechanical.

Sauti TTS supports:

• Swahili speech synthesis

• reference-audio-conditioned voice transfer

• zero-shot voice cloning

• long-text chunking

• configurable generation parameters

The model uses Vocos as the vocoder layer through the upstream F5-TTS stack.

Training on Real Swahili Speech

Sauti TTS was fine-tuned on the WaxalNLP Swahili TTS subset (swa_tts), one of the highest-quality publicly available Swahili speech corpora released to date.

📚 Dataset: https://huggingface.co/datasets/google/WaxalNLP

The dataset contains studio-quality Swahili recordings from native speakers and gave us the foundation we needed to build a system that actually sounds like Swahili.

Good speech synthesis does not begin with architecture.

It begins with data.

And speech data quality is harder than most people realize.

Training Dataset Statistics

Dataset Split

Why Audio Quality Quietly Breaks TTS Models

Even professionally recorded speech contains subtle artifacts:

• microphone handling clicks

• digital pops

• buffer underrun glitches

• resampling distortions

• spectral discontinuities

Most humans barely notice them.

Neural TTS systems do.

A speech model reproduces everything inside its training data — including invisible defects.

Train on audio with clicks, and the model learns to generate clicks.

That forced us to build a dedicated audio quality pipeline before training Sauti TTS.

Building the Audio Cleaning Pipeline

We built a multi-stage preprocessing and quality analysis pipeline for Swahili speech curation.

Stage 1 — Artifact Detection

We used mathematical analysis of the audio waveform — specifically first-derivative spike detection — to identify sudden amplitude discontinuities inconsistent with surrounding speech.

These discontinuities often correspond to:

• clicks

• pops

• corrupted sample boundaries

• recording glitches

Stage 2 — Signal Repair

Detected artifacts were repaired using smooth interpolation techniques that reconstruct damaged waveform regions while preserving surrounding acoustic continuity.

Stage 3 — High-Fidelity Resampling

Standard resampling pipelines introduced subtle boundary artifacts during 16kHz → 24kHz conversion.

We replaced the default resampling stack with a higher-fidelity pipeline designed to minimize spectral discontinuities.

Stage 4 — Quality Scoring

Every training sample was scored for:

• noisiness

• discontinuity

• clipping

• spectral coloration

• overall signal quality

Only samples above a quality threshold entered the final training corpus.

Data Preparation Pipeline

The Sauti TTS preparation stack included:

• resampling

• silence trimming

• loudness normalization

• Swahili text normalization

• metadata export compatible with F5-TTS

These preprocessing stages were critical for improving speech consistency and reducing audio artifacts during synthesis.

Teaching a Machine To Read Swahili Properly

One of the hardest problems in TTS sounds deceptively trivial:

text normalization.

Humans instantly know how to interpret:

• Ksh 5,000

• 2026

• Dkt. Amina

• Prof. Mwangi

A model does not.

Without normalization, every sentence containing numbers, currencies, abbreviations, or honorifics sounds broken.

Sauti TTS includes a full Swahili text normalization pipeline handling:

• number-to-word conversion

• currencies

• dates

• abbreviations

• honorifics

• common written conventions

For example:

When normalization works correctly, users never notice it.

That is exactly the point.

Zero-Shot Voice Cloning

Sauti TTS also supports zero-shot voice cloning.

With only a short reference recording — roughly five to fifteen seconds — the model can generate new speech in the same voice identity without additional training.

That opens the door to:

• audiobook narration

• personalized educational systems

• conversational agents

• accessibility tools

• localized voice experiences

For African languages, this matters enormously because voice identity and accent authenticity are deeply important to user trust.

Training Configuration

This release corresponds to the multi-GPU fine-tuning recipe used for Sauti TTS.

The released checkpoint package includes:

• model_last.pt

• vocab.txt

• training_config.json

• LICENSE

• THIRD_PARTY_NOTICES.md

The uploaded checkpoint corresponds to a full multi-GPU fine-tuning run and includes:

• model weights

• EMA weights

• optimizer state

• scheduler state

Checkpoint metadata reports:

update = 15350

Under the Hood

Sauti TTS is not only a model release.

It is also a systems engineering project.

The stack includes:

The goal was not simply to train a model.

The goal was to build a usable Swahili voice platform.

Beyond TTS: Building a Swahili Voice Stack

Sauti TTS is one half of a larger system.

The other half is Sauti ASR, our Swahili automatic speech recognition system.

Together, Sauti TTS and Sauti ASR create a complete Swahili voice platform:

• systems that can listen

• systems that can speak

• systems that can eventually converse naturally

That foundation enables:

• conversational AI

• educational systems

• accessibility tools

• voice search

• interactive voice systems

• healthcare interfaces

• multilingual assistants across East Africa

Sauti-Loop: Closing the Feedback Loop

We are also developing a research direction called Sauti-Loop.

The idea is simple but powerful:

ASR Errors → Targeted TTS Synthesis → Synthetic Data → ASR Retraining

Instead of treating ASR and TTS as isolated systems, they improve each other iteratively.

ASR weaknesses generate targeted synthetic speech.

That synthetic speech then improves ASR robustness.

The cycle compounds over time.

We believe this represents a promising direction for low-resource speech systems, especially for African languages where high-quality labeled speech remains limited.

Intended Use

Sauti TTS is intended for:

• Swahili TTS research

• African language speech generation

• voice interface experimentation

• conversational AI systems

• accessibility tooling

• benchmarking and reproducibility work around F5-TTS fine-tuning

Limitations

Like any speech synthesis system, Sauti TTS still has limitations.

Current limitations include:

• output quality depends heavily on reference audio quality

• quality outside Swahili has not been established

• waveform boundaries may still require cleanup during inference

• this release is optimized primarily for research workflows

We believe being explicit about limitations is important for responsible deployment and reproducible research.

Open and Built in Kenya

Sauti TTS is open-weight.

The repository and trained models are being released publicly.

• Code License: MIT

• Model License: CC-BY-SA-4.0 (aligned with WaxalNLP licensing)

This work was built entirely by MsingiAI, Kenya.

We did not have massive compute clusters or billion-dollar infrastructure.

What we had was:

• a real problem

• high-quality data

• strong architectures

• disciplined engineering

• persistence

And we believe that is enough.

Why This Matters

African languages should not exist only at the margins of AI systems.

They should exist at the center.

Not as demos.

Not as afterthoughts.

Not as “low-resource edge cases.”

But as first-class languages capable of powering modern conversational systems.

Sauti TTS is our contribution toward that future.

Try Sauti TTS

Model Weights

🤗 https://huggingface.co/msingiai/sauti-tts

Repository

https://github.com/Msingi-AI/sauti-tts

Dataset

https://huggingface.co/datasets/google/WaxalNLP

Base Architecture

F5-TTS v1 Base

Citation

@misc{sauti_tts_2026,
  title={Sauti TTS: Swahili Text-to-Speech via F5-TTS Fine-tuning on WaxalNLP},
  author={MsingiAI},
  year={2026}
}

Final Thought

The goal was never to build a system that merely pronounces Swahili words correctly.

The goal was to build a voice system that actually sounds like Swahili.

Natural rhythm.

Natural pronunciation.

Natural speech.

Because two hundred million Swahili speakers deserve AI systems that sound like them.