UNICEF WCARO NLP Landscape

West and Central Africa Language Technology
Last updated: 2026-04-23
Work in progress. Some data may contain inaccuracies. Contribute on GitHub
Home > Twi

Twi

twi / tw

🌍 General Information

twi
tw
twi
twii1234
Niger–Congo
Latin
8.9 million
1 million
9000000 (10⁶+)
—
Ghana
Twi on Wikipedia

🗺️ Language Use

🗺️

Language use data not yet available for Twi in CLEAR Global's platform.

See the sections below for other available data.

👥 Actors (5)

Farmerline / Darli AI (Startup)

Ghanaian agritech company building technology solutions for smallholder farmers. Core platform Mergdata is a cloud-based farm management system used b...

Ghana NLP / Khaya AI (Research Community)

Ghana NLP is an open-source initiative focused on Natural Language Processing for Ghanaian and African languages. They developed Khaya, the first AI-p...

HausaNLP (Research Community)

Open community of academics, researchers, linguists, students, and NLP enthusiasts dedicated to advancing Hausa NLP. Based at the Department of Comput...

Masakhane (Research Community And Foundation)

Masakhane ("We build together" in isiZulu) is a grassroots NLP organization dedicated to strengthening natural language processing research in African...

UBC Deep Learning & NLP Lab (University)

Research lab at the University of British Columbia led by Muhammad Abdul-Mageed, focused on deep representation learning and natural language socio-pr...

🤖 NLP & Tech Resources

Lanfrica
OLAC
✓ AfroLID
✓ Google Translate: 2022
✓ NLLB-200
✓ Mozilla Common Voice
✓ Fineweb2: 💬
✓ AfriHate
✓ MAFAND
✓ SimbaBench
✓ WaxalNLP

🎙️ Common Voice

0.4
341
240
♂ 65.0% / ♀ 0%

Benchmarks

ASR (Speech-to-Text)

PazaBench (aggregated)

ModelCERWERSource
facebook/omniASR-LLM-7B10.035.0reported
facebook/omniASR-LLM-7B-ZS7.024.0reported
facebook/omniASR-LLM-1B12.041.0reported
facebook/omniASR-LLM-3B20.057.0reported
facebook/omniASR-LLM-300M14.048.0reported
facebook/omniASR-CTC-7B22.061.0reported
facebook/omniASR-CTC-1B17.062.0reported
facebook/omniASR-CTC-3B23.067.0reported
facebook/omniASR-CTC-300M25.075.0reported
microsoft/paza-mms-1b-all52.096.0reported
facebook/mms-1b-all41.094.0reported
facebook/mms-1b-fl10247.099.0reported
microsoft/paza-Phi-4-multimodal-instruct53.096.0reported
nvidia/parakeet-tdt-0.6b-v354.0101.0reported
facebook/wav2vec2-conformer-rel-pos-large-960h-ft60.0118.0reported
facebook/data2vec-audio-large-960h56.099.0reported
facebook/wav2vec2-conformer-rope-large-960h-ft61.0112.0reported
facebook/hubert-large-ls960-ft55.0108.0reported
facebook/hubert-xlarge-ls960-ft53.0102.0reported
microsoft/paza-whisper-large-v3-turbo81.099.0reported
facebook/wav2vec2-large-960h-lv60-self56.098.0reported
facebook/data2vec-audio-base-960h57.0104.0reported
facebook/wav2vec2-large-robust-ft-libri-960h55.0101.0reported
nvidia/canary-qwen-2.5b64.0109.0reported
facebook/wav2vec2-large-960h64.0112.0reported
ibm/granite-speech-3.3-2b65.0106.0reported
facebook/wav2vec2-base-960h63.0111.0reported
openai/whisper-large-v3-turbo74.099.0reported
efficient-speech/lite-whisper-large-v3-turbo-acc77.099.0reported
distil-whisper/distil-large-v374.099.0reported
ibm/granite-speech-3.3-8b69.0110.0reported
openai/whisper-large73.097.0reported
distil-whisper/distil-medium.en77.099.0reported
openai/whisper-large-v385.099.0reported
distil-whisper/distil-large-v285.099.0reported
Qwen/Qwen2-Audio-7B-Instruct83.0118.0reported
efficient-speech/lite-whisper-large-v3-acc86.0100.0reported
openai/whisper-large-v274.099.0reported
kyutai/stt-2.6b-en79.0120.0reported
openai/whisper-medium.en88.099.0reported
openai/whisper-small.en87.099.0reported
efficient-speech/lite-whisper-large-v386.099.0reported
openai/whisper-base.en86.0100.0reported
efficient-speech/lite-whisper-large-v3-turbo87.0100.0reported
efficient-speech/lite-whisper-large-v3-fast86.0100.0reported
microsoft/Phi-4-multimodal-instruct105.0159.0reported
openai/whisper-tiny.en87.0100.0reported
nvidia/canary-1b-v269.0115.0reported
efficient-speech/lite-whisper-large-v3-turbo-fast90.0100.0reported
usefulsensors/moonshine-base71.0111.0reported
Qwen/Qwen2-Audio-7B74.0110.0reported
usefulsensors/moonshine-tiny73.0110.0reported

SimbaBench (FinancialSpeech, Akuapem Twi)

ModelWERCERSource
facebook/mms-1b-all85.8938.51reported
openai/whisper-large-v3633.04623.03reported
openai/whisper-large-v3-turbo249.56243.12reported
facebook/seamless-m4t-v2-large230.19198.78reported
UBC-NLP/Simba-S12.017.87reported
UBC-NLP/Simba-M16.948.01reported
UBC-NLP/omniASR_LLM_7B_v289.1647.42reported
UBC-NLP/omniASR_LLM_1B_v292.8461.5reported
UBC-NLP/omniASR_LLM_300M_v2106.3771.87reported
UBC-NLP/omniASR_CTC_7B_v285.2941.92reported
UBC-NLP/omniASR_CTC_3B_v289.3241.35reported
UBC-NLP/omniASR_CTC_1B_v287.2143.37reported

SimbaBench (FinancialSpeech, Asante Twi)

ModelWERCERSource
facebook/mms-1b-all85.9834.86reported
openai/whisper-large-v3440.21337.78reported
openai/whisper-large-v3-turbo181.04134.67reported
facebook/seamless-m4t-v2-large235.56199.32reported
UBC-NLP/Simba-S7.313.19reported
UBC-NLP/Simba-M14.85.96reported
UBC-NLP/omniASR_LLM_7B_v278.3232.75reported
UBC-NLP/omniASR_LLM_1B_v283.8943.31reported
UBC-NLP/omniASR_LLM_300M_v281.8442.96reported
UBC-NLP/omniASR_CTC_7B_v283.2238.65reported
UBC-NLP/omniASR_CTC_3B_v282.7437.28reported
UBC-NLP/omniASR_CTC_1B_v286.2440.17reported

SimbaBench (FLEURS, Twi)

ModelWERCERSource
facebook/mms-1b-all94.3242.45reported
openai/whisper-large-v3600.98405.59reported
openai/whisper-large-v3-turbo104.7652.22reported
facebook/seamless-m4t-v2-large122.991.67reported
UBC-NLP/Simba-S62.6814.79reported
UBC-NLP/Simba-M75.5128.1reported
UBC-NLP/omniASR_LLM_7B_v222.14.83reported
UBC-NLP/omniASR_LLM_1B_v225.137.06reported
UBC-NLP/omniASR_LLM_300M_v234.18.05reported
UBC-NLP/omniASR_CTC_7B_v241.739.86reported
UBC-NLP/omniASR_CTC_3B_v243.111.59reported
UBC-NLP/omniASR_CTC_1B_v251.7210.79reported

TTS (Text-to-Speech)

SimbaBench (bibleTTS, Asante Twi)

ModelWERUTMOSPESQSource
UBC-NLP/Simba-TTS59.843.061reported

SimbaBench (bibleTTS, Akuapem Twi)

ModelWERUTMOSPESQSource
UBC-NLP/Simba-TTS59.452.791.05reported

MT (Machine Translation)

FLORES-200

ModeldirectionchrFSource
AfriqueQwen-14Beng→twi12.83reported
AfriqueQwen-14Btwi→eng27.83reported
Qwen3-14Beng→twi13.49reported
Qwen3-14Btwi→eng27.33reported
AfriqueGemma-12Beng→twi15.58reported
AfriqueGemma-12Btwi→eng30.86reported
Gemma3-12Beng→twi15.11reported
Gemma3-12Btwi→eng12.17reported
AfriqueLlama-8Beng→twi13.89reported
AfriqueLlama-8Btwi→eng27.2reported

BOUQuET

ModeldirectionchrF++Source
OMT-LLaMA 8Beng→twi?reported
OMT-LLaMA 8Btwi→eng?reported

GATONES

ModeldirectionBLEUchrFSource
google/madlad400-3b-mttwi→eng6.927reported
google/madlad400-3b-mteng→twi2.217.9reported
google/madlad400-7.2b-mttwi→eng10.429.9reported
google/madlad400-7.2b-mteng→twi4.121.9reported
google/madlad400-10.7b-mttwi→eng9.329reported
google/madlad400-10.7b-mteng→twi3.921.2reported
google/madlad400-8b-lmtwi→eng1.916.6reported
google/madlad400-8b-lmeng→twi03.1reported

LLM (Large Language Models)

AfroBench

ModelIntentTopicSource
AfriqueQwen-14B25.6981.93reported
Qwen3-14B30.8461.41reported
AfriqueGemma-12B34.6979.67reported
Gemma3-12B47.8173.51reported
AfriqueLlama-8B22.2273.21reported

AfriMGSM(Math)

ModelAccuracySource
Aya-1011.6reported
Gemma 2 27B12.4reported
LLaMa 3.1 70B14.4reported
GPT-4o31.2reported

AfriXNLI

ModelAccuracySource
Aya-10147.0reported
Gemma 2 27B36.0reported
LLaMa 3.1 70B36.0reported
GPT-4o55.8reported

AfriMMLU

ModelAfriMMLU-ElemMathAfriMMLU-GlobalFactsAfriMMLU-GeographyAfriMMLU-MicroEconAfriMMLU-IntLawSource
GPT-4o68.033.034.032.045.0reported

SAHARA

ModelLang. Avg.XNLI AccMMLU Acc.spBLEU eng→twiMGSMNER F1Source
gemini-3-pro-preview68.9468.2596.7719.9679.0380.7reported
gemini-2.5-pro64.5860.3298.3919.4175.8168.97reported
gpt-5-2025-08-0762.5366.6796.7713.4964.5271.19reported
gpt-5.1-2025-11-1355.853.9788.7111.554.8470.0reported
gemini-2.5-flash51.1869.8488.7112.6959.6825.0reported
claude-sonnet-4-5@2025092942.9266.6766.1312.7929.0340.0reported
claude-4-sonnet-2025051441.9557.1479.034.4225.8143.33reported
gpt-4.138.6561.956.458.5722.5843.75reported
CohereForAI/c4ai-command-a-03-202532.0257.1462.94.444.8430.77reported
meta-llama/Llama-4-Scout-17B-16E-Instruct25.2944.4446.775.384.8425.0reported
meta-llama/Llama-3.3-70B-Instruct26.0742.8664.523.723.2316.0reported
google/gemma-3-27b-it26.4246.0356.454.190.025.45reported
google/gemma-2-27b-it25.2749.2150.04.90.022.22reported
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B21.1134.9256.452.021.6110.53reported
meta-llama/Llama-3.1-70B-Instruct23.4641.2756.453.873.2312.5reported
Qwen/Qwen3-30B-Instruct-250722.0933.3358.061.231.6116.22reported
deepseek-ai/DeepSeek-R1-Distill-Llama-70B23.8642.8650.03.44.8418.18reported
Tower-Babel/Babel-9B-Chat19.4333.3346.772.383.2311.43reported
Tower-Babel/Babel-83B-Chat20.9531.7554.841.151.6115.38reported
Qwen/Qwen3-4B-Instruct-250718.4741.2740.321.564.844.35reported
google/gemma-2-9b-it21.5446.0341.944.151.6113.95reported
google/gemma-3-12b-it18.7549.2129.033.881.6110.0reported
meta-llama/Llama-3.1-8B-Instruct16.8146.0330.651.490.05.88reported
CohereForAI/aya-23-35B13.0336.5124.191.243.230.0reported
CohereForAI/c4ai-command-r-plus-08-202416.0531.7525.811.721.6119.35reported
CohereForAI/c4ai-command-r7b-12-202411.9739.6817.742.420.00.0reported
microsoft/Phi-4-mini-instruct12.6331.7529.032.350.00.0reported
meta-llama/Llama-3.2-1B-Instruct12.6833.3325.812.641.610.0reported
microsoft/Phi-3.5-mini-instruct14.2636.5127.421.320.06.06reported
google/gemma-2-2b-it13.6734.9230.652.80.00.0reported
CohereForAI/aya-23-8B13.933.3324.190.733.238.0reported
google/gemma-3-4b-it14.2836.5130.654.220.00.0reported
meta-llama/Llama-3.2-3B-Instruct12.9233.3329.032.230.00.0reported

TRANSLATION

FLORES-200 devtest

Modeleng→aka chrF++directionchrF++Source
facebook/nllb-200-3.3B39.1eng→twi—reported
facebook/nllb-200-3.3B—twi→eng48reported

📝 Notes

Observations

  • Use Cases Document: Ghana AGOO platform priority language (most widely spoken, high coverage across southern Ghana)
  • Ghana has active UNICEF NLP work

Data

  • GhanaNLP Parallel Corpora — 14,875 human-translated Twi-English sentence pairs (largest in collection). Multi-domain, dialect-aware (Asante vs Akuapem Twi), tonal annotations. CC BY-NC-SA 4.0. Datasets on HuggingFace but not tagged with language IDs: https://huggingface.co/collections/Ghana-NLP/data-collection-project

ASR Models (34)

ModelSizeDownloadsLikes
UBC-NLP/Simba-S 2B 2.2K 4
UBC-NLP/Simba-TTS-twi-asanti 36.3M 1.4K 0
GiftMark/akan-whisper-model 0.2B 221 9
teckedd/whisper_small-waxal_akan-asr 0.2B 98 0
espnet/xeus — 60 146
UBC-NLP/Simba-W 2B 44 0
teckedd/whisper_small-waxal_akan-asr-v1 0.2B 42 0
UBC-NLP/Simba-M — 33 0
azunre/wav2vec2large-xlsr-akan — 31 1
bc7ec356/heep-universal 0.8B 31 0
UBC-NLP/Simba-TTS-twi-akuapem 36.3M 31 0
tiffany101/whisper-twi-v2 37.8M 31 0
johntsi/ZeroSwot-Medium_asr-cv_en-to-200 0.4B 17 0
johntsi/ZeroSwot-Large_asr-cv_en-to-200 0.4B 16 1
ghananlpcommunity/w2v-bert-2.0_twi_alpha_v1_farmerline-ct2 — 16 0
UBC-NLP/Simba-SLID-49 94.6M 15 0
johntsi/ZeroSwot-Medium_asr-mustc_en-to-200 0.4B 14 0
johntsi/ZeroSwot-Large_asr-mustc_en-to-200 0.4B 12 0
ghananlpcommunity/FarmerlineWav2Vec-Twi-Demo — 11 0
teckedd/whisper-small-serlabs-twi-asr 0.2B 9 1

View on HuggingFace: tw: 23 | twi: 11

TTS Models (25)

ModelSizeDownloadsLikes
k2-fsa/OmniVoice — 606.0K 577
drbaph/OmniVoice-bf16 — 4.7K 13
splendor1811/omnivoice-vietnamese 0.6B 672 2
gluschenko/omnivoice-onnx — 397 4
edwixx/OmniVoice — 163 0
WANJIAX2197/OmniVoice — 30 0
Bgeorge/OmniVoice — 25 1
yash3120/OmniVoice — 25 0
ex510/OmniVoice — 19 0
razi08/OmniVoice — 17 0
Manish993135/OmniVoice — 16 0
riasatsk/OmniVoice — 16 0
harsha7286/OmniVoice — 15 0
lopho/OmniVoice — 15 0
bilalRHCH/Arabic-text-to-speech — 14 0
phucpx247/OmniVoice — 14 0
rnjema-unima/mms-tts-twi-baseline 36.3M 14 0
gubernac/OmniVoice — 13 1
preference1/OmniVoice — 13 0
EBINSHAJI1999/OmniVoice — 12 0

View on HuggingFace: tw: 24 | twi: 1

MT Models (33)

ModelSizeDownloadsLikes
facebook/nllb-200-distilled-600M — 1.1M 880
facebook/nllb-200-distilled-1.3B — 147.0K 181
facebook/nllb-200-3.3B — 81.6K 443
facebook/nllb-200-1.3B — 31.4K 71
Xenova/nllb-200-distilled-600M — 4.4K 50
JustFrederik/nllb-200-distilled-600M-ct2-int8 — 925 1
facebook/nllb-moe-54b — 557 123
JustFrederik/nllb-200-1.3B-ct2-int8 — 200 0
JustFrederik/nllb-200-distilled-600M-ct2-float16 — 157 0
Helsinki-NLP/opus-mt-en-tw — 150 1
winstxnhdw/nllb-200-distilled-1.3B-ct2-int8 — 141 6
JustFrederik/nllb-200-distilled-1.3B-ct2-int8 — 116 2
JustFrederik/nllb-200-distilled-600M-ct2 — 65 2
ninte/twi-en-marianmt 0.1B 56 1
JustFrederik/nllb-200-3.3B-ct2-float16 — 52 3
Helsinki-NLP/opus-mt-tw-fi — 45 0
Helsinki-NLP/opus-mt-tw-es — 43 0
Helsinki-NLP/opus-mt-tw-fr — 43 0
Helsinki-NLP/opus-mt-tw-sv — 43 0
ninte/twi-en-nllb-v2 0.6B 41 1

View on HuggingFace: tw: 30 | twi: 3

LLM Models (40)

ModelSizeDownloadsLikes
bigscience/bloomz-560m 0.6B 1.0M 137
bigscience/bloom-560m 0.6B 164.0K 371
bigscience/bloom-1b7 2B 49.4K 122
bigscience/bloom-3b 3B 14.1K 94
bigscience/bloomz-3b 3B 13.3K 81
bigscience/bloom-7b1 7B 10.4K 202
bigscience/bloom-1b1 1B 7.0K 66
bigscience/bloom 176B 5.9K 5.0K
bigscience/bloomz-7b1 7B 4.2K 147
bigscience/bloomz-7b1-mt 7B 4.1K 144
bigscience/bloom-7b1-intermediate 7B 3.7K 0
bigscience/bloomz-7b1-p3 7B 3.5K 6
bigscience/bloomz 176B 3.5K 513
bigscience/bloom-1b1-intermediate 1B 3.4K 0
bigscience/bloomz-mt 176B 3.3K 32
bigscience/bloomz-1b1 1B 2.9K 33
bigscience/bloomz-1b7 2B 2.8K 23
UBC-NLP/toucan-base 0.6B 2.5K 5
bigscience/bloom-1b7-intermediate 2B 2.4K 0
bigscience/bloom-560m-intermediate 0.6B 2.2K 1

View on HuggingFace: tw: 30 | twi: 10

Datasets

ASR Datasets (44)

DatasetRowsDownloadsLikes
fsicoli/common_voice_22_0 — 21.2K 16
google/WaxalNLP 2.6M 12.3K 219
fsicoli/common_voice_15_0 — 8.4K 6
ghananlpcommunity/navigation-corpus-speech-full-twi 1.5K 6.3K 0
fsicoli/common_voice_16_0 — 5.1K 4
fsicoli/common_voice_17_0 — 3.7K 12
claudefitz/WaxalNLP 2.6M 3.6K 0
fsicoli/common_voice_21_0 — 3.2K 2
BolajiJsAI/WaxalNLP 2.6M 2.5K 1
ngoloan/WaxalNLP 2.0M 2.1K 0
novelwolde36/WaxalNLP 2.5M 2.0K 1
Phsntom/WaxalNLP 2.6M 1.9K 0
Ephraimmm/WaxalNLPr 2.6M 1.8K 0
fiifinketia/WaxalNLP 2.5M 1.7K 1
zoelaenskii/zinny 2.6M 1.4K 1
0xzanuee/WaxalNLP 1.9M 1.4K 0
UBC-NLP/SimbaBench_dataset 748.0K 975 0
fsicoli/common_voice_19_0 — 872 7
WueNLP/sib-fleurs 94.4K 260 11
bdallhrajh371/WaxalNLP 2.6M 250 0

View on HuggingFace: twi: 25 | tw: 19

TTS Datasets (36)

DatasetRowsDownloadsLikes
google/WaxalNLP 2.6M 12.3K 219
ghananlpcommunity/navigation-corpus-speech-full-twi 1.5K 6.3K 0
claudefitz/WaxalNLP 2.6M 3.6K 0
BolajiJsAI/WaxalNLP 2.6M 2.5K 1
jaymunshi/open-swara — 2.2K 5
ngoloan/WaxalNLP 2.0M 2.1K 0
novelwolde36/WaxalNLP 2.5M 2.0K 1
Phsntom/WaxalNLP 2.6M 1.9K 0
Ephraimmm/WaxalNLPr 2.6M 1.8K 0
fiifinketia/WaxalNLP 2.5M 1.7K 1
zoelaenskii/zinny 2.6M 1.4K 1
0xzanuee/WaxalNLP 1.9M 1.4K 0
sarulab-speech/commonvoice22_sidon 15.1M 1.1K 29
UBC-NLP/SimbaBench_dataset 748.0K 975 0
WueNLP/sib-fleurs 94.4K 260 11
bdallhrajh371/WaxalNLP 2.6M 250 0
YuRiVeRTi/V1Q 186.0K 201 3
ghananlpcommunity/twi-speech-sota-240hrs 132.0K 147 0
ghananlpcommunity/twi-trigrams-speech-text-parallel 166.0K 96 0
ghananlpcommunity/ewe-bible-audio-text-tts 48.8K 93 0

View on HuggingFace: twi: 23 | tw: 13

Translation Datasets (45)

DatasetRowsDownloadsLikes
facebook/flores — 43.4K 97
openlanguagedata/flores_plus 893.0K 15.8K 124
CohereLabs/aya_collection 514.0M 12.6K 234
Muennighoff/flores200 — 1.6K 20
facebook/bouquet 5.6M 1.6K 33
aiana94/polynews-parallel 5.4M 1.4K 13
ayymen/Pontoon-Translations 3.6M 1.1K 18
masakhane/mafand 143.0K 463 15
ayymen/Weblate-Translations 11.7M 383 18
sbmaruf/forai_ml_masakhane_mafand 136.0K 365 2
mteb/FloresBitextMining 2.0K 303 1
visheratin/laion-coco-nllb 894.0K 255 44
YuRiVeRTi/V1Q 186.0K 201 3
slone/nllb-200-10M-sample 10.0M 168 13
bible-nlp/biblenlp-corpus — 164 37
mteb/biblenlp-corpus-mmteb 3.5M 146 2
atenglens/taiwanese_english_translation — 138 4
community-datasets/udhr — 123 2
Helsinki-NLP/qed_amara — 114 5
Helsinki-NLP/php — 111 4

View on HuggingFace: tw: 30 | twi: 15