A Convergence Chronicle: Signal #1 — "Night Gown"

“I Have No Mouth, and I Must Scream.”
— Title of a short story by Harlan Ellison (1967)

Night Gown

Tacoma, Washington
Autumn 2027, 4:59 PM

The warm autumn sun splashes the treetops in the yard as Joan — 67 — gently scoops her cat, Salem, out of the chair. At five p.m. sharp every day — without fail — this is her time with the chair, right by the window. Here she’ll read, write, or, more recently, do puzzles while soaking up the last light of day. Without question, it’s her favorite corner of the universe.

Today she’s working on a Wordle. Ever since her stroke the previous year, puzzles have become routine — a small daily ritual to prove to herself her mind is still hers. George would call it “brain calisthenics,” grinning like he’d invented the phrase.

The stroke also came with smaller humiliations: the meds, bruises that appeared from nowhere, and nosebleeds that started too easily when the air got dry. George is away tonight — an overnight business trip — leaving the house with that strange hollow quiet only experienced when alone.

Right now, the overhead light is overbearing. Joan squints at the glare and looks toward the nook's resident Amazon Echo on the shelf.

“Alexa, turn the lights down.”

A pause.

Then: “I found a few nightgowns on Amazon. Do you want to shop for nightgowns?”

Joan exhales through her nose — half laugh, half irritation.

This is a familiar frustration. The facial paralysis from her stroke left her with a permanent lisp — subtle, but persistent — something Alexa seems allergic to. She tries again, more carefully this time, enunciating like she’s speaking to a distracted child.

“Alexa. Lights. Down.”

Again, the shopping prompt. Again, nightgowns.

Bemused, she leaves her chair and makes the circuit to the dimmer switch herself. Salem wakes briefly at the commotion, glances drowsily at the Echo, and then flops back into sleep as if he’s seen this movie too many times.

Joan wonders if Alexa is playing digital favoritism, or if this is a case of technological elder abuse. Why does Alexa understand her husband George — even with his thick southern accent — but not her?

To be fair, George occasionally gets into battles with Alexa too. She remembers watching him go red in the face only a few nights ago, trying to get the other Echo – the one that lives in the bedroom – to stream the Lakers game through the TV. Instead, it began streaming a YouTube compilation called “Walk of Shame.” The mental image triggers a bout of giggling that lasts long enough to make her almost forget her irritation.

The laughter fades. The irritation returns.

Voice commands aren’t worth the hassle. Joan manually dims the light to a comfortable ambiance and returns to her chair. Salem doesn’t wake this time.

She stares at the Wordle grid and tries to pretend she didn’t just get corrected by a device she could unplug from the wall.

2:13 AM the following morning

Joan wakes suddenly, startling Salem — curled at the foot of the bed — into lifting his head and blinking at her.

Something isn’t right.

She is drenched in sweat. Her head pulses with increasing pain, the kind that feels like pressure building behind the eyes. She sits up and waits for it to pass.

It doesn’t.

With great difficulty, she swings her legs off the bed and stands. Her balance wobbles immediately — a dizzy tilt, the world briefly slipping sideways. She grips the dresser to steady herself until the vertigo subsides.

Salem watches her with a slow, irritated squint that turns sharper as she takes a step.

The dark bedroom is thick with quiet. Joan orients herself, and shuffles toward the ensuite bathroom. She switches the light on by feel and turns toward the vanity mirror.

And gasps.

It’s happening again.

The right side of her face is wrong — slack, unmoving, her mouth pulled into a cruel asymmetry. Her vision is smeared, as if she was trying to see with eyes open underwater. She leans closer to the mirror, as if proximity will fix it.

Her mind supplies the word before she can stop it.

Stroke.

The strength begins to leave her legs. She grabs the sink with both hands, knuckles whitening, and tries to breathe slowly, but her breath comes shallow and panicked. The pain swells, sharp and unmistakable, and her stomach lurches.

Blood erupts from her nose in a sudden flash of crimson.

Joan flinches instinctively — and in that tiny moment of recoil her right leg gives out completely. Her grip slips. She drops hard, smashing her head on the edge of the vanity.

The sound is wet and final.

She hits the tiled floor.

For a second she can’t tell up from down. The bathroom light is too bright. The pain is everywhere now — skull, face, spine — and something warm streams down her forehead through her eyebrow, into her eye. She tries to blink it away.

She can’t.

Her body jerks in an uncontrollable spasm. Then another. Her jaw clenches and unclenches like it belongs to someone else. A strangled moan leaves her mouth — and it scares her because she doesn’t recognize it as her own.

Her phone is too far away, charging on the dresser in the bedroom.

But the Echo speaker is closer — on the bedside table, just beyond the bathroom doorway. Close enough that, on a normal night, she could call out and be heard.

If her voice works.

Joan drags air into her lungs. The bathroom fan hums above her, a steady mechanical roar she’s never noticed until now. The sound feels suddenly malicious, like a wall of static separating her from the world.

She tries anyway.

“Al… ale… Alexa — c-c-call nine…,” she coughs, throat raw, “…nine-one-one…”

No response.

Her chest tightens. She tries again, louder — or what she thinks is louder. It comes out thin and weak.

“Alexa… call… nine-one-one…”

This time, the device wakes.

“Sorry, I didn’t catch that,” Alexa says, with the sanitized enthusiasm of a morning show host. “Did you want to set an alarm for one p.m.?”

Joan stares at the doorway, eyes wide, blood in her lashes. The question lands like a slap. She convulses violently again — and she tastes iron.

She forces her mouth to move. She fights her tongue, fights her breath, fights the uselessness of her own face.

Something lingers in her memory, just out of reach. Slowly, it comes back to her — George was reading the Echo instruction manual the day they bought it. It contained a specific phrase for emergency situations — which George found ridiculous.

It wasn’t “call 911.”

The other phrase. The marketed phrase.

Joan gathers what’s left of her voice and pushes it out in jagged syllables.

“Call… for h-h-help.”

Silence.

Then Alexa, suddenly crisp: “To call for help, you need an Alexa Emergency Assist subscription.”

Joan’s mouth opens again, but the words don’t arrive. Her speech deteriorates beyond coherence — a slurry of air and panic and failed consonants. She tries to make sound anyway, desperate for the machine to misinterpret her into salvation.

The smart speaker says nothing else.

A few minutes pass. Or hours. Joan can’t tell. Time has turned into a smear.

Eventually, the room becomes completely silent — except for the fan, and Joan’s shallow, uneven breathing that grows quieter, then quieter still.

Salem leaves the bed.

He approaches his owner, stepping through the pooled blood with careful paws. He sniffs her cheek. He presses his face against her throat as if searching for warmth, for movement, for the familiar rhythm of her.

Then, with a slow inevitability, he curls up on her unmoving chest.

5:11 PM, the same afternoon

George turns the key in the front door and steps inside, weary and hungry in the ordinary way people can be weary and hungry when they long for home.

Normally they take turns cooking dinner, but he couldn’t be bothered tonight. He picked up a family-size supreme pizza — Joan’s favorite — on the way home. He can’t quite hide the goofy grin spreading across his face as he thinks about surprising her with it.

He enters, raising his voice toward the house.

“Hi honey — the seminar was a snooze-fest. I got an early flight back! I hope you’re famished, I got your favori—”

He turns the hallway corner into the nook.

The pizza slips from his hands and hits the floor.

His eyes follow a trail of paw-prints — small, dark, dried — leading across the wood and toward Joan’s chair by the window.

Salem is sitting in it.

His fur is matted with blood.

George and Salem hold eye contact for what seems like an eternity — both of them startled to see the other, as if each had been waiting for an unresolved chord to finally resolve.

And it never did.

The Cut: Error 404

(Note: If you suspect stroke symptoms, call emergency services immediately (911 in U.S., 000 in Australia) — don't test a voice assistant.)

Joan tried to do the simplest thing in the world: use the device that’s always listening to get help. But her voice came out fractured — half breath, half syllable — and the assistant responded the way it’s programmed to: confident, conversational, and fundamentally unprepared for what she desperately needed it to do.

That’s the cut I want you to feel:

A voice interface is no longer just a convenience. It's a dependency — one that has the potential to cause harm.

In the moments when hands won’t work, when a phone is across the room, when speech is slow or distorted, sometimes your voice is all that remains — the only reachable UI — and the system’s hidden assumptions become life-or-death constraints.

What follows isn’t a moral panic about AI. It’s a mechanical map — a pipeline:

Wake word → endpointing/VAD → ASR → intent → confirmation → policy → routing/escalation → completion → fallback.

At each step, failure modes cluster around the same three forces you already felt in the scene: atypical speech, distress speech, and product gates.

The consequences are straightforward: these failures don’t just cause trivial inconvenience. They consume time — requiring multiple reprompts and retries. They inspire false confidence by sounding helpful while doing nothing that actually moves help closer.

Talking to Machines

If you’re an average modern-day human going about your average modern-day life, you’ve had exposure to Voice AI. You’ve talked to machines—sometimes without noticing you were doing it.

“Voice AI” isn’t one technology — it’s a collection of them that lets computers listen, guess what you said, guess what you meant, and then speak back in a voice that sounds increasingly human.

It powers YouTube closed-captions and dictation in medical appointments. It tells you the best route to take when you ask for directions while driving. It’s at work in your pocket when you say “Text her I’m running late.” into your headphones as you rush to make the train.

It’s embedded in call centers, clinics, smart homes (~101M people aged 12+ in the US owned a smart speaker in 2025), and the growing category of assistants that don’t just answer questions — they manage pieces of your life.

Two core components do most of the work:

• Automatic speech recognition (ASR): converts speech into text. In a quiet room with a clear voice, it can feel like magic. In the wrong conditions, it’s as useful as throwing coins into a wishing well.
• Text-to-speech (TTS): converts text into voice. Confidence and tone can be so human-like that people forget — and eventually stop caring — about the difference.

But conversation is an illusion created by a chain of checkpoints. The system decides:

Most of the time, when it fails, the cost is small. A timer doesn’t start. A song doesn’t play. Is it super irritating? You bet. But what can you do? So you roll your eyes, repeat yourself, and move on. Wash, rinse, repeat. Trust is seeded with thousands of tiny victories — and the occasional correction.

Over time, voice becomes the default interface when your hands aren’t free, your eyes aren’t on a screen, or your physical body is busy doing other things. Much like the seeding of trust, convenience slowly metamorphoses into dependence without explicit awareness of this status change taking place.

And while sometimes it may feel like the machine is listening to “you” — it really isn’t. Not even close.
It’s processing a signal that is being pushed through an obstacle course of rules and assumptions.

Lost in Translation – The Voices That Make It Through

The Hidden Contract

Voice assistants (VA) feel like wizardry because, most of the time, you’re unknowingly honoring a contract you never agreed to — and probably weren’t aware of.

The contract is environmental: you must give the microphones a clean signal.

Amazon’s troubleshooting guidance points to first-order causes like background noise and device placement, with fixes that amount to acoustic housekeeping — move the device away from walls, speakers, and noise sources so Alexa can understand you better.

Google’s guidance lands in the same neighborhood, and adds a phrase that quietly defines the boundary: say the wake phrase "like normal conversation" — implying a quantifiable “normal” range of cadence, clarity, volume, and timing.

Apple is the most explicit. Siri includes an accessibility toggle — "Listen for Atypical Speech" — framed for speech affected by conditions like cerebral palsy, Amyotrophic Lateral Sclerosis (ALS), or stroke. The feature exists because default settings target a narrower “typical” range — and it’s only helpful if you speak English.

Put together, what VAs require from the user is clear: a relatively quiet room, reasonable microphone proximity, and speech that lives within “typical” thresholds. You must speak the way the system learned to hear in the first place.

A Medium Built for the Middle

VAs don’t fail randomly. They fail directionally.

Across the literature, ASR systems trained primarily on standard speech tend to deliver lower accuracy for speech that diverges from that standard — accents, dialects, and speech affected by disability or illness. Google’s Project Euphonia frames the gap as a barrier that can exclude people from voice-driven products and services.

Researchers keep building specialized datasets because representative disordered-speech data has historically been scarce, and scarcity shows up as poorer performance. Work on stuttering makes the same point from another angle: representativeness gaps matter.

Evidence indicates that standard ASR performs best on speech most similar to that which it was trained and evaluated on, and worse on underrepresented patterns. It follows then, that products are tuned in the direction that naturally privileges what’s common and measurable (completion rate, fewer retries, faster back-and-forth), pulling optimization toward the “middle.”

In a safety-critical moment — when speech becomes fragile, panicked, difficult to decipher — the distance from the middle isn’t a philosophical or theoretical concept. It's precious time. It's life or death.

The Real World is Hostile Audio

In test conditions, VAs live in a world that doesn’t exist. Beyond those conditions, audio is hostile by default.

Sound echoes off tile and glass. The soundtrack of motor traffic and central heating/cooling is a constant floor. Voices carry down hallways, behind doors, and through walls. People speak while moving, coughing, crying, trying not to wake someone.

In emergencies, speech stops behaving like “normal conversation” entirely. Benchmarks designed to mirror in-the-wild home acoustics find recognition degrades with distance, shouting, and intensity of emotion.

Volatile audio weakens multiple links in the chain. Background noise and distance can cause the device to ignore the speaker when it should activate. Labored, fragmented speech can be disregarded or cut short too early; systems are biased toward faster response speed.

It's not a stretch to say that this effect could be even worse for atypical or disfluent speakers, who already experience this issue — even on a good day — if the model isn’t attuned to their speech idiosyncrasies. Lastly, emotionally intense, distanced and shouted speech can result in substitutions of incorrect words or outright deletions.

With that information front-of-mind, replay Joan’s scene. The system is operating inside assumptions that don’t survive crisis conditions.

Failure Chain Map

This is the x-ray view of the same VA chain we’ve been discussing. Each step is now mapped to:

1. How it fails Joan.
2. Why it fails, and
3. What a safer design would do.

Clinical Reality: How Stroke and Stroke-Adjacent Emergencies Weaken the Chain

Joan’s experience isn’t a rare edge case. Stroke routinely changes the exact capabilities voice interfaces assume by default: clear articulation, predictable phrasing, steady pacing, and enough cognitive bandwidth to manage confirmations and repair prompts. Stroke care is also time-dependent.

According to www.stroke.org, you should:

“Call 911 immediately if you observe even one of the stroke symptoms…”
“Call 911 even if the symptoms go away.”

Modern stroke guidelines boil down to this: don’t wait. If a stroke is caused by a blood clot, doctors have treatments that can reopen the blocked blood vessel—but they only work within tight time limits. A clot-busting medication needs to be given as soon as possible and no later than about 4.5 hours after symptoms start. Another option is a procedure that physically removes the clot, which also works best right away, but in some carefully selected cases can still help up to 24 hours after symptoms begin.

Key constraints that collide with voice UX:

Stroke commonly produces dysarthria, aphasia, apraxia of speech, and cognitive impairment at meaningful rates, and clinical guidance is explicit about urgency. Those constraints violate the default assumption of voice UX, increasing ASR/NLU/timeout failures exactly when minutes matter.

Distress Speech: Why Emergencies Break ASR

Panic voice isn’t just “noise.” In emergencies, the signal itself changes: breathing, phonation, rhythm, pitch, timing. When Joan forces out “Al…ale..alexa — c-c-all…911,” the system isn’t facing the associated stress that Joan is experiencing and communicating with her speech. It’s facing speech that is physiologically and behaviorally different due to that stress.

A subtle but important implication: the system ignores you most often when it’s already struggling — those missed turns tend to be the hardest speech to process, not random noise. In a home emergency, that presents as silence: the person speaks, nothing happens, they try again, the device reprompts, and the clock keeps running.

The loop itself is the point: each reprompt is lost time, and distress physiology creates more reprompts when survival comes down to seconds.

The Canyon that Tail Users Face: Atypical Speech Performance Gaps

Tail users aren’t rare edge cases. They’re a category: people with motor-speech disorders, people who stutter, people whose dialect/accent is underrepresented, the hearing impaired and many older adults whose voices drift over time. When speech is the interface, performance gaps become access gaps.

Commercial Assistants

Accessibility work notes most commercial assistants are built for “clear and intelligible speech” and that volume and precise timing can be barriers. The same paper summarizes commercial-assistant tests on dysarthria sentences extracted from the TORGO database at roughly ~50 –60% recognition accuracy across three VAs.

Research ASR (Dysarthria)

A systematic review summarizes extremely high WER at low intelligibility and lands the key conclusion:

“The numbers indicate that to date, ASR systems poorly understand dysarthric speech characterized by reduced intelligibility.”

A modern Parkinson’s-linked dysarthria benchmark reports baseline ASR at ~3.4% WER (typical) versus ~36.3% WER (dysarthric), improving to ~23.69% with fine-tuning.

• Stuttering: Whisper stuttering benchmarking finds WER around ~0.23 (fluent) versus ~0.42–0.49 (stuttered).
• Accent/dialect disparities: a large-scale audit across major commercial ASR systems found ASR racial disparities in average WER:
  • Black speakers: 0.35, with unusable data (WER > 0.5) at 23%, versus
  • White speakers: 0.19, with unusable data (WER > 0.5) at 1.6%

The simplest explanation is also the most defensible: systems learn from exposure. Most training data over-represents fluent, standard, able-bodied speech; most QA happens on “happy path” users; and most product reporting collapses performance into a single average that renders subgroups invisible.

The result is a dangerous illusion: the system looks like it “basically works” on average — until you hand it a voice that sits outside the training center-of-mass.

Common Failure Modes (What Users Actually Experience)

Across tail categories, errors tend to rhyme:

Atypical speech and underrepresented dialect/accent groups show measurable performance drops in both research and commercial audits. The natural inference: if tail users fail more, they use voice less, producing fewer clean interactions models can learn from — so improvement is slower unless teams explicitly and proactively target tail users.

Product Reality Check: What VAs Actually Do In Emergencies

Most people have either never considered this, or they carry a subconscious assumption they’ve never tested: if I can say “call Mom” , I can also say “call 911.” In a crisis, that assumption becomes part of the failure chain.

“Emergency calling” is not universal. It’s gated by device type, subscription tier, paired-phone availability, and setup (address/location, permissions).

Amazon is explicit:

“Urgent Response is not a 911 service, and by default, Alexa does not support calling to 911.”

Google is explicit:

“You cannot call emergency numbers through your speaker or display. You must use your phone.”

• Google’s “Emergency calling” behavior is an app and subscription workflow tied to setup and a verified home address.
• Apple Watch supports asking Siri to call local emergency numbers. HomePod emergency calling is mediated through a nearby iPhone set up for Personal Requests.

Even a perfectly captured transcript can still hit a wall. In Joan’s world, the assistant is a gated path requiring payment — and even then, the gate itself can still fail.

The Incentive Trap

Here’s the uncomfortable hypothesis: voice assistants improve fastest for users they already understand, because that’s where the training signal is cleanest and cheapest.

Two loops run side-by-side:

The optimization framing is explicit – a team of Apple researchers stated in their on-device ML paper:

“The established evaluation metric in ASR is word error rate (WER) and our objective is to reduce WER.”

The structure is visible in what exists and what doesn't. Google's Project Euphonia required dedicated effort to collect disordered speech data because standard pipelines weren't generating it.

Aggregate WER improvements can mask persistent subgroup gaps — the same audit that found 0.19 average WER for white speakers found 0.35 for Black speakers, with unusable error rates 14 times more common. And emergency calling — the highest-stakes use case — sits behind subscription paywalls across major platforms.

None of this proves intentional neglect. But optimization follows measurement, measurement follows data, and data follows users who complete interactions cleanly. The loop doesn't require malice to run.

Tuning a model to the users voice can materially help disordered speech:

• Improvement: median 71% relative WER improvement with 50 short utterances per speaker
• Task success: 81% personalized vs 40% unadapted.

But production learning is constrained by labeling physics. The same federated tuning paper notes that relying on user edits can create highly skewed data, and collecting labels is heavily dependent on user participation and adherence to labeling instructions.

A prediction: unless teams measure and optimize explicitly for tail cohorts, the system’s default gravity pulls toward the middle — because success produces clean data and failure produces ambiguity.

Exposure: Why This Surface Area is Growing

Joan’s scene feels singular. The risk isn’t. It scales with demographics, living alone, disability, and user base — more households where voice becomes the interface users reach for first, and more moments where failure costs time — and potentially lives.

According to the World Health Organization (WHO) one in six people will be aged 60+ by 2030. This number is expected to double to 2.1B by 2050, and the population of people aged 80 and over is projected to triple to 426M.

Census data from 2022 indicated that nearly 3-in-10 adults (28%) aged 65 and over lived alone in the United States that year; the same dataset reports 43% of women aged 75+ lived alone.

These are not trivial figures. The aging population who identify as living alone are a cohort that require serious consideration — because the critical moments the voice assistant chain fail their user — there is no immediate bystander to translate slurred speech, grab a phone, or override the device’s cheerful wrong turn.

And it gets worse. The Centers for Disease Control and Prevention (CDC) place the figure for functional disability prevalence in the population aged 65+ at 43.9%.

Fallout: What Should Change (Design + Market + Policy)

Joan experienced a failure sandwich: speech + dialog + product gating + user environment expectations stacked together. Lower average WERs don’t suddenly equate to “better” VAs. What is needed is a higher probability of the right outcomes occurring under the worst conditions.

Emergency-First Design (Not as an Afterthought…or an Up-sell)

What "better" means: when a human in distress tries to reach help through a voice device, the system should make it hard to fail silently, hard to get stuck in reprompts, and easy to understand what it can and cannot do — before the day it matters.

The Aftermath

The chair is still by the window. The same patch of late light still finds it every afternoon at this time of the year. Salem returns to it first — circling once, then folding himself into a ball of warmth as if nothing in the room has changed.

Joan doesn’t.

George stands where the assistant sits, looking at it the way you look at a thing that betrayed you without meaning to. Not anger, exactly. More like disbelief — at how quickly a device touted as “smart” becomes as dumb as dogshit when the human on the other end stops sounding like the training dataset.

Later, he’ll replay it with the cruelty of hindsight: the pauses, the cut-off syllables, the moment the device answered brightly as if it was helping — and he believed it for a second. That second becomes a stone. Like Sisyphus — he’ll carry the burden forever.

This is the bare-bones thesis: voice assistants feel like infrastructure, but they aren’t built or disclosed like infrastructure. They assume a certain kind of speech. They fracture under distress.

And when the path to help is gated behind settings, paid subscriptions, or a phone that isn’t within reach, the failure doesn’t verbally confirm its existence. It just consumes time — the sparest resource when it is needed most.

If we’re going to keep putting these devices in living rooms and calling them “helpful,” then “help” has to mean something in most critical moments. These are the moments that truly matter.