Our new research with the GSMA shows that voice-first AI can run fully offline on low-cost hardware, down to a Raspberry Pi.

Most AI today assumes you’re online. Every question goes to a data centre and comes back. That works if you have fast, cheap, reliable internet. For millions of people across Africa, you don’t.

Many African language communities also talk more than they type. If AI is going to be useful to them, it has to listen and speak, and it has to work when the network doesn’t.

That’s the problem behind our new paper, Offline AI Modules: Voice-First Offline Architecture, Hardware Reference Stack, Quantization and Benchmarking. This paper was written by Awarri researchers- Sunday Afariogun and Odunolaoluwa Jenrola, in collaboration with Zeinab Nezami of the GSMA team, as part of the GSMA African AI Language Initiative.

What we built

We designed a voice assistant that runs entirely on one small device, with no connection needed. It works in three steps:

1. It listens. Speech recognition turns what you say into text.

2. It thinks. A compact language model works out a response.

3. It replies. Text-to-speech says the answer back.

Alongside the design, we published a parts list for two hardware setups. One is built around a Raspberry Pi 5. The other is built around an NVIDIA Jetson Orin NX, a compact edge-AI board. We also published a step-by-step method for shrinking AI models so they fit on these devices, and for testing how well they still perform.

What we found

We tested three open models (Gemma 4 E2B and Gemma 2 2B, and Qwen2.5 3B) at different levels of compression. This is the first time this full offline stack has been benchmarked end to end on both kinds of hardware.

• It fits on a Raspberry Pi. All three models ran within the Pi’s memory, using under 2.6 GB of RAM at peak. The first response arrived in about two seconds.

• It’s fast on edge hardware. On the Jetson, the best model began replying in about half a second and generated nearly 29 tokens per second.

• Compression doesn’t have to cost quality. In the Q4_K_M quant form , the best model used about a third of the memory of the uncompressed version and ran more than twice as fast. It still scored 89% on a news-classification test in English, Hausa, Igbo, Nigerian Pidgin and Yoruba.

• It runs on modest power. The Jetson peaked at 20–24 watts, with no overheating.

We were also honest about the gaps. Speech recognition for Amharic did well, but Oromo lagged badly because there’s far less training data. Hausa was the weakest of the text languages for two of the models. These gaps are where the work needs to go next.

Why offline matters

Running AI on the device isn’t just a workaround for bad connections. It changes what’s possible:

• Access in places with little or no signal

• Reliability when networks drop

• Privacy, because voice data never leaves the device

• No cloud bills, so there’s no cost every time someone asks a question

• Local experimentation on hardware that universities, start-ups and community groups can actually afford

“If AI only works with a strong connection, it leaves out the people who could benefit most. We wanted to show that useful voice AI can live on a small device, in the community, in their language.”

- Sunday Afariogun, Awarri

What’s next?

This is a foundation, not a finished product. Next, we want to:

• Cut the response time on the Raspberry Pi

• Help close the speech-data gap for under-served African languages

• Run African-built language models through the same pipeline as they’re released

• Expand our Benchmarks to cover more tasks.

  
  

At Awarri, we believe AI built for Africa should work in African languages and in African conditions. This research is one step toward voice AI that reaches people where they are, online or not.

Read the paper on arXiv.

Working on offline or African-language AI?
Get in touch by sending us an email.