Listening to the Firehose: Can Voice-First, Two-Way Audio Become a Legitimate Assistive Medium?

Thesis

Voice-first, two-way audio agents are technically ready to become a useful complement to screen-based reading for knowledge workers, but only if designers treat them as assistive, user-controlled, and hearing-safe tools—not as replacements for reading or as always-listening ambient companions.

Why this matters

Knowledge workers spend most of their waking hours in front of screens. The American Optometric Association and Deloitte Access Economics estimate that more than 104 million working-age Americans now spend over seven hours a day on screens, and that unmanaged screen time cost the United States economy roughly $151 billion in 2023. Digital eye strain is common enough that prevalence estimates range from 25% to 93% depending on the population studied, with symptoms that include dry eyes, blurred vision, headaches, and neck or back pain.

At the same time, audio is already a large part of daily life. Edison Research reports that 76% of Americans aged 12 and older listened to online audio in 2024, and 47% listened to a podcast in the last month, both record highs. Podcast advertising has grown into a multibillion-dollar market, with comedy and sports leading genres. Audio commands attention: a Lumen and Dentsu study found that audio ads, especially in podcasts, generate more attentive seconds per thousand impressions than many visual ad benchmarks.

The correlation is clear: people are saturated with screens, and they are already comfortable with audio. What is not yet clear is whether the next generation of voice interfaces—agents you can talk with, interrupt, and ask questions—will actually help people understand complex information, reduce fatigue, and protect wellbeing, or whether they will simply replace one cognitive burden with another.

The screen-first trap

For professionals who consume text-heavy information, the default interface is still a glowing rectangle. Email, documents, chat threads, dashboards, and social feeds all compete for the same visual channel. The result is not just eye strain; it is a structural attention problem. Reading allows re-reading, skimming, tab switching, and self-pacing, but it also invites interruption and visual clutter.

Reducing screen time sounds simple in theory and is hard in practice. Workflows are built around screens, collaboration tools assume visual attention, and much professional knowledge is stored in formats that are easy to scan but hard to consume while walking, cooking, or resting the eyes. The goal should not be to eliminate screens, which remain unmatched for many tasks, but to create legitimate alternatives for moments when the visual channel is overloaded or unavailable.

What listening can and cannot do

Listening and reading are not interchangeable. Jiang, Sabatini, Wang, and O’Reilly found that working memory and attention explained 16% to 25% of the variance in listening comprehension, compared with 11% to 12% for reading. One reason is pacing: readers can slow down, re-read, or skip ahead, while listeners must keep up with a transient stream. Sweller, van Merriënboer, and Paas describe this as the transient information effect: spoken content disappears and must be held in working memory, which increases cognitive load.

That does not mean listening is inferior. Mayer and Moreno’s modality effect shows that spoken narration can offload an overloaded visual channel when it is paired with well-designed visuals. For some tasks, audio is the better channel. The problem is that many audio products treat the listener as a passive recipient rather than an active participant.

A provocative finding by Jiang, Kalyuga, and Sweller illustrates the complexity: in a foreign-language learning study, more expert learners performed better with reading-only than with listening-only or read-and-listen conditions, while novices benefited from dual modalities. The lesson is not that one modality wins; it is that the right format depends on the learner, the material, and the context. Voice-first consumption should therefore be framed as a complement to reading, not a wholesale replacement.

The new audio interface

Until recently, most voice assistants were cascaded pipelines: automatic speech recognition, then a language model, then text-to-speech. The result could feel slow, turn-bound, and brittle. That is changing. Ji et al.’s 2024 survey of spoken dialogue models documents the shift toward end-to-end speech models that handle speech representation, streaming, full-duplex interaction, and turn-taking in a single system.

Kyutai’s Moshi, described by Défossez et al., is one concrete example. It is a speech-text foundation model for real-time dialogue, built with a streaming neural codec called Mimi, and it models two audio streams simultaneously: the user and the agent. The reported latency is roughly 160 milliseconds in theory and about 200 milliseconds in practice on an L4 GPU. That is fast enough to support natural turn-taking, interruptions, and backchannels.

The technical milestone is real, but it does not answer the design question. A conversation that feels natural is not automatically a conversation that improves comprehension. Naturalness can even work against usefulness if the agent interrupts at the wrong moment, supplies confident-sounding but inaccurate summaries, or nudges the user toward engagement rather than understanding.

Designing for comprehension, not just convenience

The design challenge is to use audio as a cognitive scaffold, not merely a hands-free convenience. Several principles from cognitive-load theory are directly relevant:

• Modality. Use audio when the visual channel is already overloaded or when the user cannot look at a screen.
• Segmentation. Break long material into short, navigable chunks rather than dumping a 30-minute monologue into the listener’s ears.
• Redundancy. Avoid reading the exact same text aloud while it is also on screen; the gains are smaller than people assume and the cost in attention can be high.
• Self-pacing. Let the user pause, repeat, skip, or ask for clarification. The listener must retain agency over the pace of information.

The AI interlocutor can help by summarizing, defining terms, answering questions, quizzing the listener, or marking a place to resume later. But every one of these features is also a potential interruption. The agent should wait to be asked, or should ask permission before speaking, rather than assuming that helpfulness means proactivity.

Trust, agency, and the attention economy

Proactive voice assistants create a dilemma. Zargham et al. found that users appreciate proactivity in urgent or critical situations but worry about loss of agency, intrusiveness, and social disruption. Kraus et al. showed that trust in proactive assistants is shaped by how predictable, transparent, and controllable the behavior feels. Oh et al. went further, arguing from a wizard-of-Oz study that proactive voice assistants should ask exploratory questions, incorporate feedback, seek permission for control tasks, and keep adjusting until the user explicitly says to stop.

These findings argue against the seductive idea of an always-listening ambient companion that surfaces information whenever it judges the moment right. Such a system might be technically feasible, but it risks becoming another attention parasite. Audio is intimate and harder to ignore than a notification badge. A voice that speaks without invitation can feel like an intrusion into the listener’s personal space.

There is also a curation risk. Li, Mithas, Tam, and Zhang found that algorithmic filtering on a social platform narrowed users’ interest scope over time and reduced exposure to attitude-challenging content. If audio agents become a primary interface for news, research, and professional updates, the same dynamics could apply in a channel where users have fewer visual cues to detect bias or verify claims. Designers should build in transparency about what is being selected and why, and users should be able to challenge or redirect curation easily.

Accessibility and hearing health

Voice-first design must include people who do not fit the default able-bodied, hearing user profile. That means captions and transcripts, adjustable playback speed, visual alternatives when the user wants them, and motor-free triggers such as voice commands or head gestures. It also means designing for users with attention differences, auditory processing disorders, low vision, or motor impairments.

Hearing health is a concrete, often overlooked constraint. The World Health Organization notes that 80 decibels is safe for up to 40 hours per week, while 90 decibels is safe for only about four hours per week. It recommends keeping device volume below 60%, using well-fitted noise-cancelling headphones, taking breaks, and monitoring exposure. A work-related increase in earbud use for spoken content has not been well studied over long periods, so caution is warranted. Audio agents that encourage longer or louder listening sessions could create a secondary health problem while solving a primary eye-strain problem.

A trigger-based off-screen review layer: a testable hypothesis

The most viable near-term form of voice-first professional audio may not be an ambient companion, but a trigger-based, off-screen review layer. Imagine a knowledge worker wearing earbuds while away from the desk. A priority notification arrives, or the user taps an AirPods stem, or says a wake phrase. The agent reads a short summary of a document, thread, or queue, then waits for a command: “explain that again,” “skip to the decision,” “send a canned reply,” or “save for when I’m back at my screen.” The session is short, scoped, and easy to exit.

Apple’s Announce Notifications feature already points in this direction: Siri can read time-sensitive or direct-message notifications through AirPods while the device is locked, and newer AirPods support head-gesture replies. The interaction is pull-based or notification-triggered, not ambiently proactive. The user remains the initiator.

This pattern is best treated as a testable hypothesis, not a settled product design. The open questions include:

• Does two-way audio review actually improve comprehension and retention compared with reading or one-way audio for professional content?
• What is the optimal length and interruption pattern for an audio review session?
• Does the pattern reduce total screen time, or does it merely shift attention across devices?
• How do users verify claims, detect bias, and avoid filter bubbles when the interface is voice-first?
• What are the longitudinal effects on hearing health and social presence?

Until these questions are answered, the responsible stance is to ship small, scoped experiments, measure comprehension and wellbeing outcomes, and give users clear controls over when the agent speaks and what it says.

Practical takeaway: a design checklist

For builders and designers considering voice-first professional audio, the following checklist distills the argument:

1. Start from a real screen-fatigue moment. Identify a workflow where users genuinely cannot or should not look at a screen.
2. Make the user the trigger. Prefer user-initiated or notification-initiated sessions over always-listening ambient companions.
3. Keep sessions short and scoped. A two-minute review of one decision is easier to process than a 20-minute digest.
4. Preserve self-pacing. Let users pause, repeat, skip, ask questions, and exit without friction.
5. Design for comprehension, not just hands-free consumption. Include summaries, definitions, and checkpoints rather than raw reading.
6. Build trust through transparency. Explain why content was selected, surface sources, and make curation challengeable.
7. Protect hearing health. Encourage moderate volume, breaks, and exposure awareness.
8. Include accessibility from the start. Provide captions, transcripts, visual alternatives, and motor-free controls.
9. Measure the right outcomes. Track comprehension, retention, screen-time change, user agency, and wellbeing—not just engagement minutes.
10. Frame it as a complement. Never position audio as a replacement for deep reading or complex visual work.

Limits, cautions, and open problems

This article is deliberately cautious. The evidence for screen fatigue and digital eye strain is strong; the evidence that audio adoption is driven by screen fatigue is correlational, not causal. The technical feasibility of full-duplex voice agents is well documented; the empirical research on comprehension with an AI interlocutor is sparse. The design recommendation for a trigger-based review layer is supported by HCI proactivity research, but there is little product-market evidence for this exact pattern.

Other limits include platform specificity: the AirPods example reflects one ecosystem, and Android, Pixel Buds, and other platforms may offer different affordances. The filter-bubble literature is contested and platform-specific, and audio recommender systems have been studied less than social-media feeds. Market figures such as the $151 billion screen-time cost and podcast revenue projections are tied to 2023–2024 data and will age quickly. Finally, the business model for AI-summarized or AI-conversational content layers is unsettled, with open questions around rights, attribution, and revenue sharing.

Conclusion

Voice-first, two-way audio agents are no longer science fiction. The underlying models are fast enough, audio is already a mainstream medium, and the problem of screen fatigue is real. But technical readiness is not the same as design readiness. The value of listening with an AI interlocutor depends on whether it respects cognition, agency, and hearing health.

The best near-term path is not an always-listening companion, but a modest, trigger-based, off-screen review layer that users control. It should be framed as a complement to reading, tested as a hypothesis, and judged by whether it improves comprehension and wellbeing—not merely by how natural the conversation feels.

Sources

• American Optometric Association & Deloitte Access Economics, The impact of unmanaged excessive screen time in the United States (2024). https://www.aoa.org/AOA/Documents/Eye%20Deserve%20More/Cost%20of%20Unmanaged%20Screen%20Time%20Report_FINAL.pdf
• Kaur et al., Digital Eye Strain – A Comprehensive Review (2022). https://pmc.ncbi.nlm.nih.gov/articles/PMC9434525/
• Edison Research, The Infinite Dial 2024. https://www.edisonresearch.com/the-infinite-dial-2024/
• Jiang et al., Are working memory and behavioral attention equally important for both reading and listening comprehension? (2018). https://pmc.ncbi.nlm.nih.gov/articles/PMC6096896/
• Jiang, Kalyuga & Sweller, The Curious Case of Improving Foreign Language Listening Skills by Reading rather than Listening (2018). https://doi.org/10.1007/s10648-017-9427-1
• Mayer & Moreno, Nine ways to reduce cognitive load in multimedia learning (2003). https://doi.org/10.1207/S15326985EP3801_6
• Sweller, van Merriënboer & Paas, Cognitive Architecture and Instructional Design: 20 Years Later (2019). https://doi.org/10.1007/s10648-019-09465-5
• Ji et al., WavChat: A Survey of Spoken Dialogue Models (2024). https://arxiv.org/abs/2411.13577
• Défossez et al., Moshi: a speech-text foundation model for real-time dialogue (2024). https://arxiv.org/abs/2410.00037
• Kraus et al., The Role of Trust in Proactive Conversational Assistants (2021). https://doi.org/10.1109/ACCESS.2021.3103893
• Zargham et al., Understanding Circumstances for Desirable Proactive Behaviour of Voice Assistants: The Proactivity Dilemma (2022). https://doi.org/10.1145/3543829.3543834
• Oh et al., Better to Ask Than Assume: Proactive Voice Assistants’ Communication Strategies That Respect User Agency (2024). https://doi.org/10.1145/3613904.3642193
• Li et al., Does Algorithmic Filtering Create a Filter Bubble? Evidence from Sina Weibo (2019). https://doi.org/10.5465/AMBPP.2019.14168abstract
• Dentsu & Lumen Research, Audio Attention Economy Study (2023). https://lumen-research.com/blog/lumen-audio-attention-economy/
• World Health Organization, Deafness and hearing loss fact sheet. https://www.who.int/news-room/fact-sheets/detail/deafness-and-hearing-loss
• Apple Support, Announce Notifications with Siri on AirPods or Beats. https://support.apple.com/en-ca/102536
• Apple Support, AirPods User Guide. https://support.apple.com/guide/airpods/welcome/web