Voice and conversation operations

Turning on Guinness voice

The browser captures microphone audio only after consent, readiness checking and browser permission. WebRTC sends audio to OpenAI; the local server keeps the API key out of the browser. Development stays on the Mac, but live speech uses the paid hosted API.

Put OPENAI_API_KEY in the ignored .env.local file in the project root. Never put it in chat, a public browser variable or Git. Restart the local server, open /demos/guinness, confirm eligibility and audio consent, then select Start voice. Set provider project spending controls; the five-minute browser timer is not a financial cap.

The initial direction is an original Irish male character, perceived age 50–60, calm and understated. Cedar is a candidate; accent, age and gender presentation must be judged by listening. Test the introduction, factual responses, unknowns, correction, interruption, mute and ending. Record what was actually heard. Do not claim successful live voice before a real session works.

The app currently does not save audio or captions. Provider processing is separate. An API key being present does not prove model access or account billing. Detailed connection and listening notes: live-voice.md.

Conversation and voice recovery — 3 October 2026

Arthur’s current review greeting is “Ah, hello there. I’m Arthur. Pull up a chair.” The visitor’s answer leads the conversation. Ask a specific question because it follows what they said; avoid a recurring history/craft menu, interview rhythm or trivia dump. Surprise comes from a fitting small story, observation or joke. Stronger Dublin delivery, charm and reciprocal light flirtation are authored directions, not measured voice abilities. Four original jokes sit in character data rather than one-off provider logic.

A failed reply is different from a closed call. Recoverable provider errors leave the microphone and conversation open; there are no automatic paid retries. Explicit expiry/authentication/quota failures and closed/failed transport release resources. A transient WebRTC interruption has a 15-second recovery window before ending. The five-minute limit is unchanged and shown with a countdown. Temporary connection details contain only a bounded error category and elapsed seconds, never raw provider messages, audio or transcript content. Starting another session clears these details.

When a user reports a stop, inspect the visible notice and session diagnostics before blaming the network. Record whether the cause was a reply, transport, account or time-limit issue. Tests must prove both continuity after a recoverable error and resource cleanup after a terminal one. Human listening must still confirm accent, conversational flow and real-world reliability.

Provider capacity and clear failure handling

A local development site still depends on the hosted voice provider’s capacity limits. A confirmed repeated Pre rate-limit failure led us to reduce live instruction overhead while retaining every researched fact. Repeated refusals now end the session and release the microphone, without an automatic paid retry. Smaller instructions do not prove that the project limit is adequate; fresh live testing and, if needed, account-limit inspection remain necessary. See voice operations for measured prompt size and diagnostics.

Pre’s mental-skills companion

Pre can use optional mental rehearsal, task-focused self-talk, process goals, attention routines and supportive setback review. He listens before suggesting a technique and asks whether it helped. General food education supports adequate fuel without restriction or punishment. He has no professional psychology, medical or nutrition credentials. The research scope, sources, limits and coaching cases for all four Gauntlet roles are in Pre’s coaching research notes. These are authored research-grounded conversation instructions, not fine-tuning or validated treatment. Human review and live conversation evaluation are still required; Pre is now authorised for local voice review at /demos/nike; this does not constitute production approval.

Reviewing Pre

Start at /demos/nike, accept the audio consent and select Start voice. Pre uses an original American voice direction, not a reconstruction of Steve Prefontaine. Listen for a dreamer speaking athlete-to-athlete: possibility and warmth before statistics or a plan. His sporting details should enrich the conversation when wanted. Test a dream, disappointment, banter, factual correction, interruption and reconnecting. Keep the on-page illustrative disclosure; do not open ordinary conversation with an affiliation lecture.

Ten dossiers, ten selected events and five campaign records are the current bounded research set. Nothing here proves complete sporting coverage, live scores, clinical competence or audible character quality. See Pre’s review packet for provenance and the four Gauntlet roles. Humans approve future character revisions; conversations never silently rewrite the DNA.