etaupe/tagbana-speech-pipeline

End-to-end audio data pipeline for Tagbana (Tagouana) indigenous language preservation — crowdsourced collection, peer validation, and ASR training data preparation # Tagbana Speech Pipeline An end-to-end audio data pipeline for indigenous language preservation, built to collect, validate, and prepare speech data for Automatic Speech Recognition (ASR) training — with a focus on Tagbana (Tagouana), a low-resource language with little to no existing digital speech data. This project addresses a structural problem in ASR research: low-resource languages are underserved because clean, labeled training data is expensive and difficult to collect from native speaker communities. Tagbana Speech Pipeline solves this with a mobile-first, low-bandwidth collection workflow built around infrastructure communities already use — WhatsApp — paired with a native-speaker-driven validation layer that ensures data quality before it ever reaches a model. ## Live Demos - 🎙️ **Speech Collector** → huggingface.co - ⚖️ **Speech Validator** → huggingface.co - 📊 **Dataset** → huggingface.co ## Why This Exists Most ASR research and tooling assumes the existence of large, clean, pre-labeled datasets — an assumption that excludes thousands of low-resource and indigenous languages from benefiting from modern speech technology. Tagbana Speech Pipeline is a reusable template for closing that gap: a crowdsourcing and peer-validation framework that any low-resource language community could adapt, not just a one-off dataset. ## Architecture ``` Native Speaker (WhatsApp) │ ▼ ┌─────────────────────┐ │ Speech Collector │ Mobile-optimized web app │ (collector/) │ 16kHz mono audio capture └─────────┬────────────┘ │ raw audio + metadata ▼ ┌─────────────────────┐ │ Speech Validator │ Peer-review interface │ (validator/) │ Consensus-based validation └─────────┬────────────┘ │ validated, clean audio ▼ ┌─────────────────────┐ │ Secure Data Warehouse│ └─────────┬────────────┘ │ ▼ ASR Fine-Tuning Rea …