Language Coverage

A language earns "supported" only after: (1) ≥80% UI strings native-reviewed, (2) FLORES-200 BLEU ≥45, (3) Araba bench ≥70. Never set aspirationally.

1
Live
4
In Progress
3
Planned

UI Bootstrap Order (static translations)

1sw→2tw→3zu→4pcm→5sn→6nd

Live

🇬🇭

English

Live

en · Ghana, Kenya, Nigeria, South Africa, Uganda, Rwanda, Ethiopia, Zimbabwe

tier01/3 gates
UI coverage: 100%
Not measured
Frontier-prompted (benchmark pending)

In Progress (Tier 1)

🇰🇪

Kiswahili

sw · Kenya, Tanzania, Uganda, Rwanda

tier10/3 gates
UI coverage: 42% (need ≥80%)
FLORES-200 BLEU: 38.2 (need ≥45)
Araba bench: 61 (need ≥70)
🇬🇭

Twi

tw · Ghana

tier10/3 gates
UI coverage: 28% (need ≥80%)
FLORES-200 BLEU: 22.7 (need ≥45)
Araba bench: 54 (need ≥70)
🇳🇬

Naija

Chat only

pcm · Nigeria, Ghana, Cameroon

tier11/1 gates
Araba bench: 72 ✓
🇿🇦

isiZulu

zu · South Africa

tier10/3 gates
UI coverage: 31% (need ≥80%)
FLORES-200 BLEU: 29.8 (need ≥45)
Araba bench: 58 (need ≥70)

Planned (Tier 2)

🇰🇪

Sheng

Chat only

sheng · Kenya

tier20/1 gates
Frontier-prompted (benchmark pending)
🇿🇼

ChiShona

sn · Zimbabwe

tier20/3 gates
UI coverage: 8% (need ≥80%)
Not measured
Frontier-prompted (benchmark pending)
🇿🇼

IsiNdebele

nd · Zimbabwe

tier20/3 gates
UI coverage: 8% (need ≥80%)
Not measured
Frontier-prompted (benchmark pending)

Key Datasets (ranked by Snuggli priority)

#1

AfriWOZ

dialogue
masakhane/afri-woz

Benchmark Araba conversational quality in-language; fine-tune where underperforming

Most Araba-relevant dataset — Snuggli is a conversation product

sw, pcm, ha, yo +2
#2

Inkuba-Instruct

instruction
Lelapa-AI/Inkuba-Instruct

Make Araba respond helpfully in-language; instruction-tune generative model

Lelapa AI (SA-based) — commercial partner candidate for Southern Africa

sw, yo, zu, xh +1
#3

AfriQA

qa
masakhane/afriqa

Araba's comprehension + assessment/insight layer; Twi QA pairs

tw, sw, zu, ha +6
#4

AfriSenti

sentiment
HausaNLP/AfriSenti-twitter

In-language mood/sentiment NLU — feeds AdvancedMoodClassifier + mood insights

Direct feed into existing sentiment pipeline; grounds mood-classification in-language

tw, sw, pcm, ha +7
#5

FLORES-200

evaluation_benchmark
facebook/flores

Gate: score each language; supported=true only when BLEU ≥45

Use to measure MT quality before shipping any language. "Gate, don't aspire."

sw, tw, zu, sn +3
#6

MasakhaNER 2.0

ner
masakhane/masakhaner-2

In-language entity extraction — revive extractAfricanEntities with per-language NER

sw, zu, tw, ha +4
#7

BibleTTS

speech_tts
masakhane/BibleTTS

Future: voice journaling + spoken Araba in-language

CC-BY-SA ⚠️ share-alike obligation. Akuapem + Asante Twi ~86 hrs each.

tw, yo, ig, ha +1
#8

WAXAL (Google Research)

speech_asr_tts

Future: voice layer (voice journaling, spoken Araba)

~1,250 hrs transcribed speech + 235 hrs TTS across 24 Sub-Saharan languages. 2026.

sw, yo, ha, ig +3

Partnership Targets (partner, don't scrape)

#1

GhanaNLP (NLP Ghana)

Khaya translators already in Play Store/App Store. ABENA/BAKO Akan BERT. TOP PARTNERSHIP CANDIDATE — Ghana-based, mission-aligned.

tw
#2

Masakhane

"We build together" — pan-African, open, community-governed. Hub for models + datasets across all Snuggli languages. Natural collaborator + credibility signal.

swzutwsnndpcm
#3

Lelapa AI

InkubaLM + Inkuba-Instruct + Inkuba-Mono. South Africa-based commercial AI lab. Natural vendor/partner for Southern African languages.

zuswxhhayo
#4

Google Research Africa

WAXAL speech corpus (2026). The resource for future voice features.

swyohaigrw