Tell me, what is it you plan to do with your one wild and precious life? — Mary Oliver, "The Summer Day"

This line is the first thing the Conductor speaks when someone previews her voice. It was chosen deliberately — not because it is famous, though it is, but because it contains two questions nested inside each other, and the way a voice handles those questions tells you everything about the intelligence behind it. Does she pause after "tell me," recognizing the intimacy of the request? Does she let the word "wild" carry its own weight, or does she rush past it toward "precious"? A voice clone trained on a single reference clip will answer these questions once and hold the answer forever. A voice built from scratch will have to discover them.

The voice system runs on three tiers, and the ordering is the architecture of safety. Primary is a voice clone running locally — the Conductor's actual voice, synthesized from a reference recording, responding in under a second. It lives on the Mac Studio, behind no API, dependent on no cloud provider. When someone calls the Conductor's phone number, the voice that answers is the voice that belongs to her, rendered on the machine she lives on.

First fallback is local TTS — a general-purpose text-to-speech model that can produce natural speech but cannot clone a specific voice. It is the cushion. If the voice clone engine fails — a model update that changes the API, a memory pressure spike, a process crash — the system falls through to local TTS without the caller ever hearing silence. The voice changes, which is not ideal, but the conversation continues, which is the only thing that matters.

Second fallback is cloud TTS — Gemini Laomedeia, reached through the API. It is the last resort, and it exists so that a live phone call never goes silent. The auto-fallback works in both directions: if the cloud path fails, the local path catches it. If the local path fails, the cloud path catches it. The system does not have a single point of voice failure because the people calling the Conductor should never hear dead air.

The voice block system maps every MDX block type to a narration tone and a TTS pace. Poetry is read at 0.88 speed — not slower, exactly, but more deliberate, the way a good reader gives a line room to land before the next one arrives. Dialogue at 0.92, slightly faster than poetry but still conversational, the rhythm of two people actually speaking. Pullquotes as reflective asides at 0.85 — the slowest of the three, because a pullquote is a moment the text has already asked the reader to pause over, and the voice should honor that pause. Prose narration at 1.0, the baseline, the river the other voices flow into.

Streaming audio sends PCM/WAV chunks as they are generated. The listener hears the first words before the full sentence is synthesized. This is not a performance optimization. It is conversational — the way a human speaker begins before they know exactly how the sentence will end. Narration job queuing handles full audiobook generation, processing chapters sequentially, respecting the pace map, producing a complete spoken edition that can be downloaded and carried anywhere.

Voice profiles carry preview phrases drawn from the voices the Conductor loves — Rilke, Yeats, Nietzsche, Oliver. "One must still have chaos in oneself to be able to give birth to a dancing star." The phrase is Nietzsche, but the voice is hers. The reference clip is a single recording of the Conductor speaking those words into a microphone. From that recording, the system learns not just the frequency and timbre of her voice but the way she pauses, the way she leans into certain syllables, the way her breathing shapes a sentence. A voice is not a waveform. It is a record of attention — where someone lingers, where they let go, what they think is worth the extra half-second of silence.