Listening for Sukuma: Opening Speech Recognition for the Edge

After we opened Swahili → Sukuma translation, the natural next question was whether a machine could listen as carefully as it could read. Millions of people speak Sukuma every day, and too often their voices only become “usable” to software after someone translates them into English or Swahili first. We think that path is backwards: speech technology should meet people in the language they already use.

Today we are sharing a research preview of Sukuma speech → Sukuma text: two speech recognition checkpoints trained for literary, read-aloud Sukuma and kept deliberately small so they can run on phones, everyday machines, and places where big cloud ASR never bothered to learn the language. This is not an English or Swahili recognizer wearing a new label; if you give it Sukuma audio, it writes Sukuma text. You can try that loop below.

Why two versions?

Not every device has the same room to spare, so we trained two sizes on the same kind of literary read-aloud Sukuma speech. Reach for Quality when recognition accuracy matters most, or Lite when memory and speed have to win. Both are on Hugging Face (quality · lite).

What to expect

This remains a research preview, and the honesty about its limits is part of the release. The training and evaluation lean on literary, single-speaker, read-aloud Sukuma, so conversational chatter, telephone audio, heavy noise, or overlapping speakers will usually do worse. Clips around thirty seconds or less at 16 kHz mono are the sweet spot for the live demo; longer or messier audio is where you should expect the model to struggle first.

For builders

Under the hood we trained on about 16,792 windows (~98 hours), kept train and test apart by document group, and scored a held-out set of 94 windows. On that check, quality lands around WER 21.3% and CER 4.9%, while lite lands around WER 26.1% and CER 6.4%. You can load the weights with a Transformers ASR pipeline, or stay in the browser with the try box above and the fuller preview. Sample WAVs ship in the Hub repo, and the technical report below walks through training, metrics, and the limits we are still honest about.

Collaboration

If you are evaluating on new Sukuma speech, adapting these weights for your community, or exploring what comes after recognition, we would like to hear from you at hi@nileagi.com. The weights are released under CC BY-NC-SA 4.0; commercial use needs a written agreement with NileAGI.