Models

Open models

Research-preview weights on Hugging Face. Small models for African languages and edge settings: Sukuma speech recognition, text-to-speech, and Swahili→Sukuma machine translation.

Sukuma Speech → Text

Two Whisper-based operating points for literary Sukuma speech recognition (document-group splits). Research preview scores on read-aloud single-speaker audio, not conversational or multi-speaker ASR.

16,792

Train windows

21.3%

Quality WER

26.1%

Lite WER

Sukuma Text → Speech

Two operating points for literary Sukuma text-to-speech in macron orthography. Research preview: input must be Sukuma text; for Swahili → speech, translate with MT first.

Swahili → Sukuma Machine Translation

Two open operating points on the same literary Swahili-Sukuma bitext (31,102 sentence pairs; document- group splits). Variants of small sequence-to-sequence models, not frontier LLMs, built for edge and remote devices. Scores are research-preview, not production translation.

31,102

Train pairs

42.6

Quality chrF2

34.8

Lite chrF2

Checkpoints

TTS · Quality

Sukuma Text → Speech (quality)

nileagi-suk-tts

Sukuma TTS for literary macron orthography (ā ē ī ō ū). Research preview; prefer Sukuma text, not English or raw Swahili. For Swahili → speech, translate with MT first.

View on Hugging Face
TTS · Lite

Sukuma Text → Speech (lite)

nileagi-suk-tts-lite

Same Sukuma TTS recipe as quality, lighter footprint for tighter edge and remote deployments.

View on Hugging Face
STT · Quality

Sukuma Speech → Text (quality)

nileagi-suk-stt

Whisper-based Sukuma ASR for literary read-aloud speech (WER 21.3%, CER 4.9% on the document-group test). Prefer 16 kHz mono; not English or Swahili ASR.

View on Hugging Face
STT · Lite

Sukuma Speech → Text (lite)

nileagi-suk-stt-lite

Same Sukuma ASR recipe as quality, lighter footprint (WER 26.1%, CER 6.4%). Aimed at tighter edge and remote deployments.

View on Hugging Face
MT · Quality · ~1.2B

Swahili → Sukuma Machine Translation (quality)

nileagi-suk-mt

Best held-out chrF2 on the literary bitext (chrF2 42.6, BLEU 20.0 on the document-group test set). One-way encoder-decoder; force start token suk_Latn with source swh_Latn. Default choice when quality matters and the device can hold a compact MT checkpoint.

View on Hugging Face
MT · Lite · ~418M

Swahili → Sukuma Machine Translation (lite)

nileagi-suk-mt-lite

Same table and recipe as quality, lighter footprint (chrF2 34.8, BLEU 12.7). Aimed at tighter edge and remote deployments where parameter count and memory dominate.

View on Hugging Face

Collection and org

The nileagi-suk collection groups Sukuma dataset and related small STT, TTS, and MT models. More NileAGI research models live on the organization page.

Hugging Face

Browse the Sukuma collection or the full NileAGI organization.

Related