nileagi/nileagi-sw-tts

Swahili TTS

Authors: Zephania and Isack Odero

Swahili text-to-speech at 24 kHz. Input is Swahili text; output is speech in a female narration voice. The voice follows a reserved prompt shipped as reference.wav. That prompt is not a scored clip and is not a training speaker.

This is not Sukuma TTS. For Sukuma speech, translate with nileagi/nileagi-suk-mt first, then synthesize with nileagi/nileagi-suk-tts.

Companion recognizer	nileagi/nileagi-sw-stt
Collection	nileagi/nileagi-sw

Demo

File	Text
samples/sample_01.wav	Habari za asubuhi. Karibu kwenye sauti hii.

Digits are spoken as words. 2500 becomes elfu mbili na mia tano. A paragraph that does not fit in one generation is split on sentences, then clauses, then words, and joined.

Summary

Task	Text-to-speech
Language	Swahili (sw)
Direction	Swahili text → Swahili speech
Sample rate	24 kHz mono
Voice	Reserved female prompt (6.7 s)
Continue-trained on that prompt	No
Long text	Split, then joined. Up to 4,000 characters in the local service
Primary check	CER of a readback, plus listening

Evaluation

Read back with nileagi/nileagi-sw-stt. The prompt clip is not in the score.

Set	n	WER	CER
Unseen short lines	5	2.2%	0.5%
Paragraph (64 s, 563 characters)	1	9.5%	5.4%

How to use

After access is approved:

HTTP API (port 8093):

curl -s -H 'Content-Type: application/json' \
  -d '{"text":"Habari za asubuhi. Salio lako ni 2500."}' \
  http://127.0.0.1:8093/v1/speech -o speech.wav

The local service applies the text frontend and the long-form split. Prefer that path for paragraphs.

Intended use

• Research synthesis of Swahili narration and short service lines
• Paragraphs that must stay intelligible after a split
• Non-commercial teaching demos

Out of scope

• Commercial products, hosted paid APIs, or advertising
• Sukuma or English as input
• A scripture clip as the voice prompt
• Training on recognizer guesses
• One unsplit generation of a long paragraph

Limitations

• A noisy or mis-transcribed prompt will sound wrong
• One generation holds about fifteen seconds of new speech beside the prompt
• A join can be audible if the cut falls mid-phrase
• Number expansion is a heuristic

Related

• Recognizer: nileagi/nileagi-sw-stt
• Sukuma TTS: nileagi/nileagi-suk-tts
• Technical report: TECHNICAL_REPORT.tex

Access

Weights are gated under CC BY-NC 4.0.

1. Open Request access at the top of this page and describe your intended use.
2. After approval, authenticate (hf auth login or HF_TOKEN) and download.

Attribute NileAGI when you use the packaged release. The non-commercial term is inherited from the pretrained speech model and the published Swahili adaptation used as the starting checkpoint. Commercial licensing: hi@nileagi.com.

License

Weights: CC BY-NC 4.0.
See LICENSE and NOTICE.md.

Citation

@misc{nileagi-sw-tts-2026,
  title        = {Swahili text-to-speech},
  author       = {Zephania and Isack Odero},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/nileagi/nileagi-sw-tts},
  note         = {NileAGI}
}