Building a voice-based customer support assistant that works in Shona, Ndebele, and code-switched Zimbabwean English.
Status: Phase 1 — retrieval backend (in progress)
Voice assistants in Zimbabwe serve English speakers well and everyone else badly. ASR for Shona is thinly researched; for Ndebele, barely at all. This project builds a working support assistant end to end, then tackles the speech layer for the languages people actually speak.
Speech → ASR → retrieval (pgvector) → LLM → TTS → Speech
- Phase 1 — RAG support chatbot (FastAPI, pgvector, Ollama)
- Phase 2 — WhatsApp channel
- Phase 3 — English voice loop (Whisper → RAG → TTS)
- Phase 4 — Shona/Ndebele ASR: dataset, baselines, fine-tuning
Python 3.11 · PyTorch 2.13 · FastAPI · PostgreSQL 17 + pgvector · Ollama (Qwen2.5) · faster-whisper · HuggingFace transformers
Build notes and benchmarks in docs/.
Tennyson Zvaita · Harare, Zimbabwe