Gate full-duplex speech models on reliability, not latency alone

A self-hosted full-duplex speech model can deliver near-real-time inference while still failing the reliability gate through hallucinations. Treat latency, semantic faithfulness, and operating cost as independent adoption criteria. For PersonaPlex, prompt changes and temperature/top-k tuning did not eliminate hallucination in the observed trial, despite fast inference on an A100-backed Modal service. This is a single bounded experiment, not a universal model verdict; fine-tuning and controlled configuration validation could change the result.

Applies when: evaluating open-source conversational speech models for production voice agents. Keywords: PersonaPlex full duplex speech hallucination latency Modal A100 tuning