nileagi/nileagi-sw-stt

Swahili STT

Authors: Zephania and Isack Odero

Swahili speech recognition. Input is Swahili audio; output is Swahili text.

This is not Sukuma ASR. For Sukuma speech, use nileagi/nileagi-suk-stt. For Swahili text → Sukuma text, use nileagi/nileagi-suk-mt.

Companion voice	nileagi/nileagi-sw-tts
Collection	nileagi/nileagi-sw

Demo

One short clip (24 kHz). The reference line is the text that was spoken.

File	Reference (Swahili)
samples/sample_01.wav	Habari za asubuhi. Karibu kwenye sauti hii.

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
import soundfile as sf

repo = "nileagi/nileagi-sw-stt"
processor = AutoProcessor.from_pretrained(repo)
model = AutoModelForSpeechSeq2Seq.from_pretrained(repo)
audio, sr = sf.read("samples/sample_01.wav")

Prefer 16 kHz mono. Files longer than 25 seconds should be split; the NileAGI service on port 8094 does that.

Summary

Task	Automatic speech recognition
Language	Swahili (sw)
Direction	Swahili speech → Swahili text
Selection	Internal bake-off on held-out local clips
Continue-trained on local audio	No
Primary metrics	WER, CER

Evaluation

Held-out clips whose text matches the audio. The companion voice prompt is excluded. Scripture slices are excluded because their stored text does not match the audio.

Source	n	WER	CER
Conversation	24	13.1%	2.9%
Narration	8	26.5%	5.3%

A 64 second paragraph from the companion voice is read back at WER 9.5% and CER 5.4%.

How to use

HTTP API (port 8094):

curl -s -F file=@clip.wav http://127.0.0.1:8094/v1/transcribe

Library:

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor

processor = AutoProcessor.from_pretrained("nileagi/nileagi-sw-stt")
model = AutoModelForSpeechSeq2Seq.from_pretrained("nileagi/nileagi-sw-stt")

Intended use

• Research transcription of clear Swahili speech
• First hop of Swahili speech → Swahili text → Sukuma text
• Offline demos and teaching materials

Out of scope

• Sukuma or English ASR
• Legal, medical, or emergency transcripts
• Presenting a hypothesis as a human record without review

Limitations

• The selection set is 24 conversational clips plus 8 narration clips. It is not a public benchmark.
• Local scripture audio was not scored.
• A cut at 25 seconds can drop a word. Overlap decoding is not in this release.
• Rare names and spoken numbers may be respelled.

Related

• Voice: nileagi/nileagi-sw-tts
• Sukuma STT: nileagi/nileagi-suk-stt
• Swahili → Sukuma text: nileagi/nileagi-suk-mt
• Technical report: TECHNICAL_REPORT.tex

License

Weights are redistributed under the MIT License that accompanied the selected public checkpoint. NileAGI evaluated and packaged the checkpoint. That MIT License, not a NileAGI share-alike licence, governs the weights.

See LICENSE and NOTICE.md.

Citation

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