Africa-Voice-AI/voicera-africa

# VoicERA Africa For the end-to-end streaming VoxCPM, RAG, and tool-calling demo, see HELLO_TRACTOR_MVP.md. Africa's open voice AI platform: build, configure, and talk to **Swahili voice agents** — in the browser or over the phone. Real-time Sauti speech models (Msingi AI) on GPU, LLM-powered personas per agent, and a full management dashboard. **Highlights** - 🎙️ **Talk to Agent** — every agent card in the dashboard opens a hands-free voice conversation (browser VAD turn-taking, barge-in, live typing, streamed per-sentence speech) plus text chat. No phone number needed. - 🗣️ **Sauti Swahili speech stack** on Modal GPUs: `msingiai/sauti-asr` (STT) and `msingiai/sauti-tts-swahili-voxcpm` — a full supervised fine-tune of VoxCPM2 on a ~500-hour Swahili corpus, **trained on the CINECA Leonardo supercomputer** (4× A100-64GB; the fine-tuning guide lives in the separate documents repository (`voicera-docs`)) — plain-Swahili CER 0.007, zero-shot voice cloning (female/male reference voices), and Swahili number/currency normalization. - ⚡ **Streaming pipeline** like modern voice modes: ASR → LLM token deltas → per-sentence TTS, pipelined so the agent starts speaking while it is still thinking. See WEB_AGENT_API.md. - 📞 **Africa's Talking telephony** — Record/Play conversation loop for real Kenyan/Tanzanian/Ugandan numbers (`voice_2_voice_server/api/africastalking.py`), alongside the original Vobiz streaming pipeline and a Jambonz bridge. - ☁️ **One-click cloud deploy**: render.yaml blueprint (backend + frontend + voice server on Render) + MongoDB Atlas + Modal GPUs. **Documentation map** | Doc | Covers | |---|---| | WEB_AGENT_API.md | HTTP API to converse with agents (streaming NDJSON, chat, voice, TTS) | | MODAL_DEPLOYMENT.md | GPU speech services on Modal: deploy, voices, secrets, telephony wire-up | | render.yaml | Render blueprint + the env vars each service needs (comments inside) | | VOIP_SETUP.md | Legacy VoIP/Vobiz setup | | This file | Local development, Docker …