Souffleur gives an AI voice conversation a small visual outline: the current topic, a short line recording its decision so far, and a few words about the previous turn. Earlier topics remain nearby, with more detail available through taps. The research prototype explores how much of a conversation can fit on a screen someone intends to consult briefly.
That is a question about representation. A transcript preserves the wording and sequence of exchanges; a topic overview foregrounds subjects and their recorded state. Understanding Souffleur means looking at how it chooses those labels, what remains visible and how a reader gets back to the words behind them.
Anton Wolter, William Benjamin and Niklas Elmqvist describe the system in Souffleur: A Glanceable Display for Mobile Voice Conversations. Version 1 was submitted to arXiv on 5 November 2026. Journal publication and peer-review status remained unconfirmed when checked on 7 November. Here, “glanceable” describes the interface’s design intention.
The current topic gets a card; earlier topics fold away
The display described in section 4.4 puts the active topic in a large, fixed-height card. It contains the topic name, its decision so far, a gist of the previous turn when that turn concerned the same topic, and a limited portion of the words currently being spoken. That gist is a summary of one turn, rather than the conversation as a whole.
Reserving the card’s height prevents a long spoken turn from pushing earlier topics out of position. Those topics receive slots when first left, alternating between the two sides of the display. The sides carry no speaker meaning.
The newest earlier topics appear as bubbles containing a name and decision. Older ones shrink to name-only pills, then clusters of four that can be opened in place. This keeps the overview compact, while making some information less immediately visible: a folded topic’s decision requires expansion.
Tapping a topic opens what each side said about it; another tap reaches the underlying words. This is progressive disclosure: an overview first, with detail available when the reader asks for it. The implementation also describes a stored word ledger in section 4.1. Keeping those words provides a route back from the labels, though speech recognition still mediates what is recorded.
A separate bottom status line tells the reader whether the system is listening, thinking, looking something up, speaking or waiting. It serves a different purpose from the decision line attached to a topic.
Short labels follow explicit generation rules
Souffleur uses several language-model calls to construct its overview. The rules in section 4.3 specify tight text budgets.
The topic titler receives the most recent 900 conversation characters, the current title and up to 16 earlier titles. It can keep the current title, select an earlier one to indicate a return, or create a new title of two to four words. It runs after a user turn adds at least 20 characters, and every 30 seconds while new speech exists.
A separate digest runs after 2.5 seconds of silence. It reads the last 80 lines and produces a decision line of at most six words per topic. It is instructed to retain concrete times, places and numbers, record only what the person agreed to, exclude unaccepted suggestions and mark unclear items as unclear.
When a new turn begins, the preceding turn is compressed into a gist of at most eight words. While that gist is pending, the display uses the last question or the last sentence with fillers removed, shown more lightly.
These are implementation instructions. They do not establish how accurately the system recognises a topic change or distinguishes an agreement from a suggestion. The calls run asynchronously, so speech can continue while the display catches up.
Speaker colour measures words, not the origin of an idea
Orange identifies the person and blue the agent. A topic receives one speaker’s label when that speaker supplied at least 65% of its words; otherwise it receives a split marker.
That rule approximates verbal contribution. It cannot establish intellectual authorship. A person might repeat an agent’s suggestion or spend many words developing it. An agent might restate a person’s idea at length. In either case, word share and idea origin can diverge.
Readers interpreting the colour therefore need to know what was counted. Recovering who first proposed something requires inspecting the conversation itself.
Compression can hide the qualification that matters
The design raises a practical tradeoff: how much wording must a reader reopen to understand a compact label?
A two-to-four-word title has little room for qualifications. A six-word decision may preserve a time and place while leaving out the reasons, alternatives or uncertainty surrounding an agreement. An eight-word gist necessarily offers less wording than the turn it represents. Grouping speech by topic also raises a question about recovering changes over time or connections between subjects.
These are possible consequences of the representation, not observed extraction failures. The underlying words may remain available even when their context is absent from the overview.
Progressive disclosure makes the route back to detail especially consequential. A short label can provide an entry point, but the reader still has to recognise when it needs checking and open the relevant exchange.
The formative work explains how the design emerged
The authors describe four months of building and using the system, mostly through their own autobiographical research-through-design, supplemented by a handful of external participants over multiple weeks.
The most frequently cited external participant was a relative of the first author: a native German speaker who completed twelve sessions over nine days. The paper says quotations were translated from German and explicitly treats those sessions as formative because of the relationship.
This material helps explain the choices the authors made. Its autobiographical basis and family involvement can influence what is noticed and reported, so it provides design input rather than independent validation. It is also separate from the paper’s later 48-person comparison.
The reported study leaves divided attention an open question
In the abstract and introduction, the authors report a 48-person between-subjects comparison with a conventional transcript, finding no reliable differences in following or resuming short voice conversations, effort, attention or recall. Their introduction says the study could detect only large differences and involved no concurrent task.
The paper’s abstract and introduction report these findings; we could not independently check the detailed study methods, results, tables or limitations. That access limit confines the study discussion here to those attributed introductory statements.
Actual use alongside another task leaves specific questions to investigate. Does a brief look preserve enough qualification to resume the intended thought? Can someone reopen the underlying words and correct their interpretation without losing their place? How do they follow a returning topic, a delayed label or an older topic whose decision has folded out of view?
