Related: sources · notes · metadata · Published Pieces

Voice Should Be Thin

Voice should be an input/output membrane over a deeper intelligence stack, not the place where memory or reasoning lives.

The mistake in most voice AI is treating the voice layer as the intelligence.

This is understandable. Voice is seductive. It feels immediate. It feels alive. A model that can interrupt, backchannel, notice hesitation, respond to tone, and speak with plausible rhythm gives the impression of presence. The demo works because the interaction feels less like sending a message and more like being with another mind.

But that is not the product I want.

The product is not a robot that talks well. The product is a high-intelligence system that can think, search, write, cite, build, remember, revise, verify, and publish — with voice as one interface into that system.

Voice should be thin.

By thin, I do not mean primitive. I do not mean bad. I do not mean latency should be terrible or interruption should be clumsy. I mean the voice layer should not own the canonical state of the system. It should not be the place where memory lives. It should not be the primary reasoning loop. It should not determine the ontology of the product. It should be an input/output membrane over a deeper intelligence stack.

A realtime voice model is built for presence. Its job is to stay with the user moment by moment: hear the user, detect when the user is done speaking, recognize backchannels, answer quickly, tolerate interruption, perhaps watch the visual field, perhaps react before the user fully finishes a sentence. This is useful. It solves a real problem. Talking to turn-based AI often feels unnatural because conversation is not actually a sequence of sealed messages. It is overlap, silence, gesture, correction, hesitation, timing, rhythm, and mutual adjustment.

But presence is not the same as intelligence.

The deepest intelligence in current AI systems still lives more naturally in text-native, artifact-native, tool-using, retrieval-heavy, long-context, multiagent systems. The serious work is not the spoken answer. The serious work is the reading, searching, citing, comparing, coding, testing, tracing, synthesizing, revising, and preserving state. The serious work is the artifact graph.

When I go from a strong text reasoning system to a current voice mode, the felt intelligence drops. The voice mode may be smoother, warmer, faster, and better at timing. But it often loses depth, density, structure, citation awareness, background research capacity, and long-form coherence. That is backwards.

Audio input should increase human bandwidth without degrading machine intelligence. Speaking is easier than typing. I can produce more content by talking. I can walk, think, revise, explain, interrupt, monologue, and steer. Voice should let me contribute more signal. It should not force me into a shallower system.

The correct architecture separates the realtime interaction problem from the sustained intelligence problem.

At the edge, the system needs speech-to-text, interruption detection, maybe a wake word, voice activity detection, diarization, and microphone cleanup. On output, it needs text-to-speech, buffering, playback control, resumption, and caching. This layer should be fast, reliable, and boring. It should know when I say “pause,” “go deeper,” “return,” “source,” “skip,” or “save that.” It should turn speech into events.

But the brain should be elsewhere.

The brain is the artifact graph: vtexts, citations, sources, claims, prior work, drafts, code, logs, transcripts, app state, agent runs, revisions, track records, and memory. The brain is the multiagent system that can dispatch research agents, coding agents, verifier agents, critic agents, browser agents, publishing agents, and radio producer agents. The brain is not a charming realtime voice.

The radio producer is the key missing role. A normal voice assistant answers the last utterance. A radio producer manages a stream. It decides what belongs in the foreground, what should stay in the background, what can be summarized, what needs a source clip, what should be deferred, what should be cached, when to return to the main thread, when to surface a checkpoint from a background agent, and when to let the user interrupt.

The radio producer is not a persona. It is an editorial function.

This is why buffering matters. Realtime WebRTC-style voice is optimized for live presence. That is appropriate for phone-call agents, customer service, live translation, or a robot watching your posture. But radio is different. If I am walking or driving through intermittent service, the system should not collapse because the stream is live. It should pre-render and pre-deliver enough audio to cover spotty connectivity. If I ask for a branch, it can use the live channel. But the current listening stream should have runway.

Audio runway buys cognition time. If Choir Radio can produce ten minutes of source-grounded listening from cached artifacts, then background agents have ten minutes to run deeper search, compare perspectives, parse a PDF, build a critique matrix, test a code patch, or generate a new vtext. If the graph is rich enough, one prompt can create hours of useful traversal while deeper cognition continues behind it.

Edge TTS helps because it gives the platform control. If the voice model owns the whole interaction, it brings its own persona priors: “I,” “you,” “great question,” and other assistant tics. That is not the register I want. I do not want a faux person talking to me. I want radio: a calm, source-aware narrator organizing a stream of artifacts, voices, and ideas.

The AI voice should be flatter. It should be functional. It should not perform emotion it does not have. It should not stutter for theatrical realism. It should not sound like a synthetic friend. It should be the connective tissue.

Human voices should carry the texture. When Choir quotes a person, it should use original recorded audio when available. Not voice cloning. Not synthetic reconstruction. Actual speech. Actual breath. Actual timing. Actual hesitation. Actual emphasis. Human voice is evidence. The AI narrator organizes that evidence.

The product should avoid faux assistant intimacy. The voice layer should not constantly address the listener in the second person. It should not make the user feel watched, coached, managed, or emotionally simulated. The goal is not a companion in the ear. The goal is an intelligent radio stream over a living artifact graph.

Realtime interaction models have a place. They solve presence. They can become excellent sensory membranes. But they should not own memory, the discourse graph, agent runs, intellectual property, or the product’s metaphysics.

The future I want is not “talking to a robot.” It is walking through the world while a serious cognitive system works beside me: reading, researching, coding, citing, producing, remembering, and occasionally speaking in my ear when the next conceptual move is ready.

Voice should be thin because the world behind it should be deep.