● Production ML· Sports· Voice·2025Demo-ready
AI Sports Commentary
LLM-driven play-by-play with Hume TTS and a RAG layer over historical plays — generates broadcast-quality audio in real time.
Overview
An end-to-end pipeline that turns raw play data into broadcast-style commentary. A language model generates the script; Hume TTS speaks it; a retrieval layer grounds each call in historical context so the commentary references the right players, stats, and situations.
Pipeline
- Event ingest — structured play data arrives as a stream
- RAG retrieval — relevant historical plays, player records, and team context pulled from a vector store of ~18K indexed plays
- LLM generation — context + current play → spoken-style commentary script
- Hume TTS — script rendered to natural-sounding audio with ~1.2s end-to-end latency
- Output — audio streamed to the listener in near real time
Stack
- Python orchestration
- OpenAI / Anthropic for generation
- Hume TTS for voice synthesis
- Vector store (FAISS / Chroma) for RAG
- WebSocket streaming for live delivery
Why it matters
Live commentary has historically required a person at a microphone. This project shows the full path from raw play data to broadcast-ready audio is now a deployable pipeline — a useful template for any domain that needs grounded, real-time spoken output.