Voice AI is one of the fastest-moving model categories in 2026 — clinical scribes, real-time customer service, transcription at scale. But the ownership decision isn't unique to voice. Any team relying on third-party APIs inherits someone else's accuracy ceiling, their latency floor, and their pricing curve. The teams pulling ahead are the ones running their own models and fine-tuning for their domain. 

We use the voice stack as the teaching vehicle because it exercises every hard trade-off at once: latency, accuracy, streaming, and inference cost. The patterns transfer directly to LLM fine-tuning, embedding models, and multimodal workloads.  

What You’ll Learn:  

• The full NVIDIA voice model stack — cascaded and unified architectures, and when to choose each  

• A repeatable fine-tuning evaluation loop (benchmark → adapt → validate) that works for any open-weight model, not just ASR  

• Deployment and runtime optimisation on Amazon EKS with G6 (L4 GPU) instances using Triton and ONNX  

• How to reason about the cost, latency, and accuracy trade-offs of self-hosting vs. API 

What You’ll Leave With: 

• A mental model of model ownership — architecture selection through to runtime optimisation — demonstrated through voice but directly transferable to LLM and embedding workloads  
• A production-ready ASR deployment on AWS Containers  
• Confidence to apply the same deploy → tune → optimise → scale pattern to TTS, LLM, or multimodal serving 

Featuring: A real-world case study from Heidi Health — production ASR fine-tuned for Australian clinical speech on AWS.