Sauti ASR: Speech Recognition for Swahili
For a long time, I kept asking myself a simple question:
Why does speech technology still struggle so much with African languages, especially in real conversations?
Not demos.
Not benchmark clips.
Not carefully cleaned datasets.
Real conversations.
The kind where someone switches between Swahili and English mid-sentence. The kind recorded on noisy phones. The kind with accents, interruptions, names, abbreviations, and imperfect audio.
That question eventually became Sauti ASR.
At MsingiAI, we believe language technology for Africa has to be built differently. Not because African languages are “impossible,” but because many systems are optimized for environments that do not reflect how people here actually speak.
Most speech systems look impressive until they meet:
- code-switching
- long-form recordings
- conversational drift
- domain-specific vocabulary
- real-world noise
That is where things start breaking.
And honestly, building Sauti ASR humbled us very quickly.
The Moment Things Became Real
When we first started experimenting with Swahili ASR, the early results looked encouraging on short audio.
Benchmarks were improving.
WER was dropping.
Inference was stable.
Then we tested long conversational recordings.
Everything changed.
You begin realizing that speech recognition is not only about language understanding. It is also about memory, continuity, context handling, and robustness under imperfect conditions.
A model can look strong on short benchmark clips and still collapse during a 7-minute clinic conversation.
That was probably one of the biggest lessons for me personally.
The best benchmark model is not always the best real-world model.
Why We Released Two Models Instead of Pretending One Solves Everything
One thing I did not want us to do was fake certainty.
So instead of releasing one “perfect” model and pretending we solved Swahili ASR, we released two tracks publicly:
- Track A → our strongest production-facing baseline
- Track B → our Omnilingual conversational ASR research preview
Track A gave us stronger benchmark performance and cleaner deployment behavior.
Track B was more experimental, but it started showing something interesting during testing: sometimes weaker benchmark models behaved more naturally on difficult conversational audio.
That mattered to me.
Because research should not only reward what scores best on a leaderboard. It should also reward what survives reality.
The Hardest Part Was Not Training
Ironically, one of the hardest parts was not training the models.
It was building the operational system around them.
We used Modal heavily for:
- training
- evaluation
- checkpoint management
- deployment
- publishing directly to Hugging Face
That infrastructure work changed how we think about shipping AI systems at MsingiAI.
A lot of people focus only on models.
But reproducibility matters.
Deployment matters.
Release workflows matter.
If your research cannot be shipped repeatedly and reliably, it becomes fragile very quickly.
The Problem We Still Have Not Solved
Code-switching.
Swahili-English conversational speech is still one of the hardest problems we face.
People do not speak in perfectly separated languages in East Africa. Conversations flow naturally between Swahili and English, often inside the same sentence.
That creates difficult failure modes:
- phonetic drift
- English corruption
- contextual instability
- substitution errors
- broken continuity across long audio chunks
And honestly, I think current ASR systems globally still underestimate how hard multilingual conversational speech really is.
Especially in African contexts.
Why This Project Matters to Me
Sauti ASR is not just another AI project for me.
It represents something bigger that we believe at MsingiAI:
African AI systems should not only exist at the research-paper level. They should exist as deployable, inspectable, usable systems built around the realities of African users.
That means:
- real deployments
- public checkpoints
- reproducible workflows
- honest evaluations
- acknowledging limitations openly
I think the future of AI in Africa will belong to teams willing to do difficult infrastructure work, difficult data work, and difficult evaluation work — not only branding.
What Comes Next
We still have a long way to go.
We want to improve:
- conversational robustness
- long-context handling
- clinical vocabulary
- code-switched transcription
- streaming inference
- real-time deployment quality
But I am excited because Sauti ASR is no longer just an idea.
It is a real system.
It has benchmarks.
It has deployments.
It has public models.
It has real-world failures.
And most importantly, it has a direction.
That matters more than hype.
Final Thought
One thing this project taught me is that building AI for African languages requires patience.
You cannot shortcut your way through conversational complexity.
You cannot benchmark your way out of real-world audio.
And you cannot build useful systems without exposing them to reality early.
Sauti ASR is still improving.
But for us at MsingiAI, this is an important step toward something much bigger:
building language technology that actually understands how Africa speaks.
