SLMming Down Latency: How NVIDIA’s First On-Device Small Language Model Makes Digital Humans More Lifelike