AMAX details on-prem real-time voice AI with NVIDIA Riva

AMAX has published a technical blog post describing its work building a real-time on-prem voice AI pipeline with NVIDIA Riva.

• The article outlines a multi-phase prototype journey: Linux and Windows development environments, a Python command-line streaming tool, a Flask and JavaScript browser prototype, and a PyWebview desktop client for microphone and loopback audio capture.
• It also describes the Riva streaming architecture, including 16 kHz PCM audio, gRPC bidirectional streaming, interim and final transcript handling, and experimentation with NVIDIA Maxine for acoustic echo cancellation and noise removal.
• The post presents performance observations such as stable streaming, the need for threaded audio ingestion, and the impact of chunk size on latency; it also mentions deployment via Docker and Kubernetes using the Riva Speech Skills Helm chart.