How NIO Cut Radio Production Costs by 80% with AI Voice Cloning

Business Background

In the increasingly competitive smart electric vehicle market, NIO focuses on in‑car interactive experience. NIO Radio is a vehicle‑mounted audio community offering music, news, entertainment, and user‑generated content.

Business Pain Points

Traditional program production involves separate script preparation, host reading, and approvals, causing long cycles.

High manpower cost: host reading accounts for over 50% of production cost, and city‑specific content increases labor.

Solution and Optimization

NIO introduced a TTS‑VC (Text‑to‑Speech with Voice Cloning) framework to replace manual reading and automate script generation with large‑language‑model‑driven news crawlers. The process is split into two parallel stages: voice generation using TTS‑VC and script generation.

Key technical advantages:

Few‑shot training reduces required data.

Low parameter count lowers compute and hardware requirements.

Controllable generation allows correction of bad cases without full model fine‑tuning.

Strong base models selected after extensive testing.

Production workflow was modularized and templated, enabling independent voice generation, script generation, and quality checks. An artificial‑evaluation system assesses accuracy, fluency, naturalness, and timbre similarity using metrics such as loss and PESQ.

Summary and Review

The AI‑driven approach cut per‑city daily labor cost by about ¥450, saving over ¥4 million annually across 27 cities, reduced production time from several hours to under 30 minutes (≈80% efficiency gain), and required only a single A800 GPU for inference. The solution is highly reusable and scalable to new cities and program types.