Private Full-Duplex Voice Agent Runtime Deployed on DGX Spark

GenAI Protos deployed PersonaPlex-7B as a private speech-to-speech runtime for real-time, bidirectional voice interactions. The implementation focuses on low-latency voice exchange, persona control, voice variants, and runtime access through a direct WebSocket protocol.

Unlike standard HTTP model endpoints, PersonaPlex uses a custom WebSocket interface. GenAI Protos treated that as an explicit architecture constraint rather than forcing the model into the wrong API shape.

The result is a private voice-agent runtime pattern for interactive assistants, companions, game characters, onboarding flows, and other real-time voice experiences.

Use case

Private full-duplex speech-to-speech runtime for real-time voice agents.

Deployment style

On-demand GPU runtime on NVIDIA DGX Spark.

Core model capability

PersonaPlex-7B speech-to-speech model with persona prompts and 18 voice variants.

Primary interface

Direct WebSocket protocol for real-time audio interaction.

Resource profile

Approximately 14 GB VRAM and on-demand startup with warm-up requirements.

Technical outcome

Deployed as part of the GenAI Protos DGX Spark model catalog and private gateway pattern.

Real-time voice agents require a different deployment model than text chat or batch transcription. They need continuous audio exchange, persona consistency, low-latency response behavior, and protocol handling that supports streaming voice frames. The challenge was to package a non-HTTP, WebSocket-based speech-to-speech model into a private operating pattern that engineering teams can evaluate and integrate cleanly.

• Respect the native WebSocket protocol instead of forcing the model into a standard chat API.
• Support real-time bidirectional audio exchange for interactive voice experiences.
• Expose persona and voice control in a repeatable runtime pattern.
• Document audio format, warm-up behavior, runtime footprint, and integration requirements.
• Keep the model discoverable inside the wider private model catalog.

GenAI Protos packaged PersonaPlex-7B as a private full-duplex speech-to-speech runtime with a clear separation between discovery and real-time interaction.

• On-demand PersonaPlex-7B runtime on DGX Spark.
• Direct WebSocket integration pattern for real-time voice exchange.
• Persona control through system prompts.
• Support for 18 distinct voice variants according to the model reference.
• Catalog registration for discoverability while documenting that the model does not expose a native OpenAI HTTP route.

The architecture separates application voice capture, WebSocket runtime, persona control, and voice output.

Voice application

Captures microphone audio or sends text/audio frames for interactive voice experiences.

WebSocket session

Maintains the real-time bidirectional connection needed for full-duplex interaction.

Persona layer

Applies system prompt and voice selection to control assistant identity and speaking style.

Speech-to-speech runtime

Runs PersonaPlex-7B inference on DGX Spark.

Audio output

Returns speech audio frames back to the application for playback.

Catalog layer

Lists the model in the private catalog for discoverability and documentation.

Runtime packaging

PersonaPlex was deployed as a private GPU-backed speech-to-speech runtime.

Protocol handling

The model was documented as WebSocket-based rather than incorrectly positioned as an HTTP OpenAI endpoint.

Persona control

System prompt and voice variant controls were captured as part of the integration pattern.

Operational readiness

Warm-up timing, model footprint, audio format, and direct connection behavior were documented.

Catalog integration

The model was added for discovery while preserving the native protocol requirements.

The implementation created a reusable private runtime for real-time voice agents where the native interaction model is full-duplex audio rather than request-response text.

Document the native protocol before designing the application interface. Separate catalog discovery from runtime interaction when needed. Expose persona and voice configuration as first-class integration controls. Capture audio format and chunk requirements for application developers. Use direct real-time protocols when conversational latency matters more than gateway uniformity.