AfriPunctCap: Pan-African Punctuation & Capitalization Model
A fast, lightweight multilingual model fine-tuned on 314 African languages to restore capitalization (Title Case, ALL CAPS) and punctuation (commas, sentence-ending periods, question marks).
Part of the AfriSpeech open-source speech initiative.
🌍 Language Coverage & Generalizability
- Directly Trained on 314 African Languages: Spans Niger-Congo, Afroasiatic, Nilo-Saharan, and Khoisan language families from AfriSpeech/afri2universal.
- Zero-Shot Generalizability to Other African Languages: Because grammatical clause boundaries, breath pauses, appositives, and conjunctions follow shared typological and phonetic patterns across related African language families, the model exhibits robust zero-shot generalizability to hundreds of other African languages and local dialects not explicitly in the training set.
🎙️ Why This Model is Essential for African Text-to-Speech (TTS)
Across African languages, writers conventionally omit commas, despite native speakers making clear, audible pauses in normal speech:
- Written Orthography Habit: Commas are routinely skipped before conjunctions (like ne in Akan, àti in Yoruba, na in Swahili), complementizers, appositives, and compound titles.
- The TTS Bottleneck: For Text-to-Speech (TTS) models, punctuation is the single most critical signal for breath pauses and prosodic rhythm. Without commas, TTS engines rush through clauses unnaturally without breathing.
- Acoustic Grounding: Fine-tuned on speech where pauses (>=350ms) were acoustically aligned with Meta MMS Forced Aligner, this model systematically compensates for commas conventionally ignored in writing, matching natural spoken rhythm.
Highlights
- Backbone:
Davlan/afro-xlmr-mini(12 layers, 384 hidden, 117M params). - Fast CPU Inference:
- INT8 Quantized ONNX: ~15–25 ms per sentence (118 MB model).
- FP16 GPU / Safetensors: ~225 MB compact half-precision weights.
- Zero Hallucination Guarantee: Non-autoregressive token classification preserves all input words in their original order.
Model Formats Available
model_quantized.onnx(118 MB, INT8 quantized for ultra-fast CPU inference)model_fp16.safetensors(225 MB, compact FP16 weights for GPU)model.safetensors(449 MB)model.onnx(449 MB)pytorch_model.bin(449 MB)- Complete SentencePiece tokenizer files
GitHub Repository
Code, data pipelines, and CLI tool: AfriSpeech/afri-punct-cap
- Downloads last month
- -