Spoken output is a stream you can only stop wholesale, so objecting to one sentence means killing the whole answer and starting the ask again.
← Back to patterns
Visual & Generative
EmergingRead-Along
A Playhead for Generated Speech
Spoken output is bound word by word to the visible text, which makes the answer addressable in time: click a word to seek to it, drag a phrase to quote it, and stop the assistant anchored to a span rather than to a moment. Every other span pattern here — Citation Trail, AI Provenance, Bias Flag — annotates a static artifact the reader scans at their own pace; this is the only one with a clock. The binding itself is mature (Speechify, ElevenLabs Reader, Apple Books); the half that matters, interrupting at a phrase rather than at a timestamp, ships nowhere yet.
Metaphor
Playhead
Interaction
Click-to-Seek, Phrase Select, Anchored Interrupt
Use case
Voice Assistants, Audio Overviews, Read-Aloud Documents, Meeting Recaps
Control mechanisms
Word-Level Binding · Scrub Bar · Quote-This-Span · Interrupt-Here
Reference examples
ElevenLabs Reader · Speechify · NotebookLM Audio Overview
Interactive Example
The live side-by-side is the comparison.
This prototype already contains the contrast between chat and the pattern response. It stands as the primary comparison surface for this page.
A stop button is as precise as the hand that presses it.
Forty-six words of speech, and one wrong sentence in the middle. Objecting means hitting stop — which arrives 620 ms late, two and a half words on, by which time the voice has moved into the fourth sentence. The model receives one fact: you stopped it. Which claim you disputed was never part of the message.
Bind each word to its timestamp and the same objection becomes “revise(span=35–35)” instead of “stop(t=9.37s)”. Everything before 00:08.50 stays accepted; the playhead parks at the start of the disputed phrase, which is exactly where a correction has to begin. A span is authored after the fact, at leisure, and is exact.
The binding itself is old — Speechify has highlighted words since 2016 and Apple Books for a decade. What none of them do is let the highlight carry an action back. Karaoke is the substrate, not the pattern: the point is that once output has a clock, it becomes addressable, and anything addressable can be quoted, sought, or argued with in place.
Speech without a clock can only be stopped. Speech with one can be corrected.
DOPE Lens
How this pattern improves AI interaction quality.
Observability
The spoken word is the highlighted word, with its exact timestamp on screen, so what the system is saying and where it is are never inferred from audio alone
Directability
Click any word to move the playhead there, or select a phrase and stop the assistant on that span — correction lands where the error is, not where the clock happened to be
Predictability
Every word carries a start and end time before playback, so seeking is exact and the same span is addressable twice
Explainability
An interrupt reports the span it caught — the sentence, not the millisecond — so the assistant can say what it thinks you objected to
Fit
Where it earns its place.
Avoid
when
when
The answer is short enough to hear out; a playhead over two sentences is ceremony.
In the wild
Where this pattern already ships.
The pattern is shipping but still maturing across products. Entries below show how different teams have approached it; expect variation as the space settles.
Reference implementations
ElevenLabs Reader · Speechify · NotebookLM Audio Overview
Detailed write-ups for each implementation are in progress.
Research Frame
Why this pattern exists, and where it breaks.
Bind the speech to the visible text word by word, so the answer becomes addressable in time — seek to a word, quote a phrase, and interrupt anchored to a span.
A transcript has no clock: it can show what was said but not where the system is in saying it, so there is nothing to point at while it speaks.
Stopping at karaoke is the trap — word highlighting alone is a decade-old reading feature, and without seek, quote and span-anchored interrupt the binding buys nothing.
Common questions
What people ask about this pattern.
When should I avoid the Read-Along pattern?
+
The answer is short enough to hear out; a playhead over two sentences is ceremony.
What problem does Read-Along solve?
+
Spoken output is a stream you can only stop wholesale, so objecting to one sentence means killing the whole answer and starting the ask again.
Why is chat the wrong fit for this?
+
A transcript has no clock: it can show what was said but not where the system is in saying it, so there is nothing to point at while it speaks.
Related Patterns
Curated connections, not just neighbors.
Hand-picked relationships — what this pattern pairs with in practice, what it replaces, what it builds on, and what builds on it.
Often paired with
Machine State, Rendered
Machine State, Rendered shows whether it is safe to speak; Read-Along shows which word the voice is on.
Alternative to
Select & Transform
Select & Transform points at static text to rewrite it; Read-Along points at a spoken word to seek, quote or interrupt there.
Extends
Mid-Stream Steering
Read-Along takes Mid-Stream Steering to spoken answers, anchoring the correction to the disputed phrase rather than a timestamp.