What full-duplex voice means in a conversation

Many voice systems connect speech recognition, a language model, and speech synthesis in a sequence. That pipeline can work well, but each handoff may add delay and can make interruptions awkward. A full-duplex model can receive user audio and generate its own audio concurrently, allowing more flexible turn-taking, pauses, and backchannels.

NVIDIA's PersonaPlex research describes a speech-to-speech model that combines text prompts for role conditioning with audio prompts for voice conditioning. The technical change is in how the voice interaction is generated; it is not evidence that the model has a stable inner personality or remembers earlier sessions.

Sources

Role and voice are separate controls

A text prompt can guide a conversational role while a speech sample conditions voice characteristics. Separating those inputs gives developers more control over who the system is asked to play and how its speech sounds. Voice imitation still calls for careful consent and clear disclosure; a voice sample should not be treated as permission to impersonate its speaker.

PersonaPlex's authors report results on Full-Duplex-Bench and an expanded customer-service benchmark. The expansion adds service scenarios beyond a generic assistant role, but the results remain the authors' evaluation of their model and test setup—not a ranking of AI companions or proof of performance in open-ended roleplay.

Sources

Better turn-taking does not solve story continuity

A voice agent can sound more responsive and still lose a character detail, misread a scene, or fail to carry a relationship across sessions. Speech timing, response quality, and long-term continuity are separate evaluation problems. A demo focused on interruptions tells you little about whether a model can sustain a character over a week of conversations.

The same separation helps users compare formats. If speaking aloud matters most, test latency and interruption behavior. If collaborative fiction matters most, test whether the system leaves room for your character, preserves agreed story facts, and accepts corrections. The two experiences can overlap, but one should not be used as a proxy for the other.

Sources

A small checklist for evaluating a voice companion

Use the same short scene in each system you compare. Speak over the response once, pause mid-sentence, correct a character fact, and ask the model to repeat a detail from earlier in the exchange. Note whether it handles the interruption, follows the role prompt, and distinguishes what it heard from what it inferred. This is a personal comparison, not a benchmark unless you publish a reproducible protocol and results.

For a different visual direction in real-time dialogue, our Gemini Live Avatar article explains Google's enterprise announcement and its access limits. Jess uses private Discord DMs rather than a voice-first interface; the product overview is linked below.

Sources and further reading

Research and product details were reviewed September 30, 2026 where applicable. Vendor policies, plans, and features can change; verify current details with the linked first-party source.