AfriSpeech/afri-punct-cap

AfriPunctCap: Pan-African Punctuation & Capitalization Restoration Model

A fast, lightweight multilingual model fine-tuned on 3.3 Million multi-sentence passages across 751 African languages and dialects to restore commas, sentence-ending periods, and capitalization in unpunctuated ASR transcripts and text.

Part of the AfriSpeech open-source speech initiative.

📊 Precision Comparison: INT8 Quantized (Default CPU) vs. FP16 (GPU)

We evaluated the exact same held-out test sets using the dynamic quantized CPU engine (INT8 ONNX) and the full precision GPU model (bfloat16 / FP16):

Language	Language Family	INT8 Comma F1 (Default)	FP16 Comma F1	INT8 Period F1	FP16 Period F1	Accuracy Retained
Swahili (swh)	Niger-Congo (Bantu)	0.641	0.649	0.953	0.953	98.8%
Zulu (zul)	Niger-Congo (Bantu)	0.609	0.628	0.860	0.886	97.0%
Hausa (hau)	Afroasiatic (Chadic)	0.579	0.611	0.899	0.908	94.8%
Yoruba (yor)	Niger-Congo (Defoid)	0.562	0.503	0.847	0.842	100.0%
Akan / Twi (aka)	Niger-Congo (Kwa)	0.540	0.543	0.899	0.870	99.4%
Ewe (ewe)	Niger-Congo (Gbe)	0.561	0.597	0.711	0.747	94.0%
Lingala (lin)	Niger-Congo (Bantu)	0.448	0.438	0.846	0.843	100.0%
Kinyarwanda (kin)	Niger-Congo (Bantu)	0.496	0.504	0.836	0.871	98.4%
Wolof (wol)	Niger-Congo (Senegambian)	0.568	0.570	0.752	0.734	99.6%
Somali (som)	Afroasiatic (Cushitic)	0.468	0.438	0.798	0.790	100.0%
Average (Core)	Pan-African Benchmark	0.547	0.548	0.840	0.844	99.8% Retained

Key Benchmark Takeaways:

1. INT8 is the Recommended Default: INT8 dynamic quantization retains 99.8% of the full FP16 accuracy while cutting memory footprint in half to 278 MB.
2. Sub-30ms CPU Latency: Single-sentence inference runs in 15–25 ms on standard commodity CPUs without needing a GPU.
3. Zero Language Tags Required: Operates in pure text mode across all languages without requiring language specification.

---

🎙️ Restoring Omitted Punctuation for African Text-to-Speech (TTS)

Across African languages, writers conventionally omit commas, despite native speakers making clear, audible pauses in normal speech:

• Written Habits: 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.
• Syntactic Restoration: Fine-tuned on 3.3 Million multi-sentence passages, this model systematically compensates for commas conventionally ignored in writing, matching natural spoken rhythm.

---

Model Formats Available

• onnx/model_quantized.onnx (278 MB, Default INT8 quantized for ultra-fast CPU inference)
• onnx/model.onnx (1.1 GB, full precision ONNX)
• model_fp16.safetensors (540 MB, Optional half-precision weights for GPU)
• model.safetensors & pytorch_model.bin (Full precision PyTorch)
• Complete SentencePiece tokenizer files

Live Interactive Demo

Try the model online: AfriSpeech/afri-punct-cap-demo

GitHub Repository

Code, pipelines, and data preparation: AfriSpeech/afri-punct-cap