Most technology deployed in dementia care treats the communication layer as secondary, the feature is the thing, and how it speaks is an afterthought. KindredMind inverts that. The communication layer is the feature. Every technical decision the system makes serves one goal: a call that leaves the person with dementia feeling calm, recognised, and heard.
What follows is an honest, complete account of the standards the companion is held to.
How the voice itself is built
Before any of the conversational standards apply, there is the voice. Your loved one needs to recognise it from the first second. KindredMind builds your voice model exclusively from short interview sessions with our setup specialist, Sarah. Three sessions of about seven minutes each happen during onboarding. Your voice model is ready after Session 1 and deepens with Sessions 2 and 3. From the Voice tab in Settings, you can check your current voice quality, preview your voice model at any time, and add more sessions through Advanced Training whenever you want to keep refining it.
Real conversations between the companion and your loved one are never used to train the voice. Your voice model is trained on your interview sessions only. This is a deliberate boundary, both for clinical reasons (the voice model should sound like the calm caregiver, not like a person who is unwell) and for privacy reasons. Your loved one's voice is never cloned or modeled. Only your voice is used to build the companion.
How we protect your voice
Your voice is yours. We treat the recordings you give us during onboarding as biometric information, with the protection that implies. They are encrypted at rest, used only to train your voice model, and never sold, licensed, or made available to any party other than KindredMind and the technology providers we use to deliver your companion.
Before we build your voice model, you sign a clear consent in your KindredMind dashboard. The consent confirms that the voice is yours, that you understand what KindredMind will do with it, and that you agree not to use your voice model to impersonate anyone else or deceive any person. This last point matters. Voice cloning is a powerful technology, and we take its potential for misuse seriously. The consent makes our shared boundaries explicit.
You can pause or delete your voice model at any time from the Voice tab in your dashboard. Deletion is permanent within thirty days. We retain the consent record itself, including the exact text you agreed to, the date and time, and the device you used, for compliance and audit purposes. The voice model itself is gone.
If you would like to read the full consent text before signing up, it is reproduced in our Privacy Policy and shown in full during onboarding before any voice training begins.
Opening the call
The first fifteen seconds of a call are the most important. For a person living with dementia, an unfamiliar opening can provoke immediate confusion or anxiety. KindredMind's opening protocol is designed to prevent that.
Standard 1
Identify clearly and immediately, with the right time of day
The companion states its name. The caregiver's name, whose voice it carries. At the very start of the call. It does not assume the person will recognise the voice without context. It says who it is, and if relevant, the simple reason for calling. The first phrase is also tailored to the local time of day where the person is, so a 7 am call opens with "good morning" and a 6 pm call opens with "lovely to hear from you this evening." A small detail that quietly anchors the person in the present moment. One or two sentences. Nothing else until the person has had time to settle.
Standard 2
Speaker detection at the start
The phone may be held by the person with dementia, but the first voice on the line is not always theirs. A spouse, an adult child, a paid carer, or facility staff often picks up. The companion listens for this. If the first voice is clearly not the patient, it gives a brief polite greeting and asks for them by name, instead of launching into the personal opening intended for someone else. This protects the person with dementia from confusion when family members handle the phone for them.
Standard 3
Check in before proceeding
Before moving to any subject or question, the companion asks whether now is a good time to talk. This is not a formality. A person with dementia may have just woken, may be mid-activity, or may be distressed from something that happened moments before. Checking in honours that. It also gives the person an immediate sense of agency over the conversation. It is their call, not something being done to them.
Standard 4
Use their name, and keep using it
The companion uses the person's name, including any preferred nickname the caregiver has specified, naturally throughout the call. Hearing one's own name is orienting. It confirms that the caller knows them, that this is not a random call. Names are used at natural pauses, not inserted mechanically.
Pacing and processing time
One of the most common errors in dementia communication is treating silence as a problem to solve. Cognitive processing in dementia takes longer. When a question is asked and no answer comes immediately, the instinct is to rephrase, repeat, or fill the space. That instinct is usually wrong. It adds noise on top of a system that is already working hard.
Standard 5
One message, one topic
The companion delivers one thought at a time. It does not combine topics into a single sentence. It does not introduce a new subject before the person has responded to the current one. Each sentence is complete and self-contained. Key words are given slight natural emphasis, not exaggerated, not over-articulated, just present.
Standard 6
One question at a time
The companion never stacks questions. It asks one, then waits. Even if the first question seems to go unanswered, it does not add a second. It waits, and then either gently restates the original or moves on naturally.
Standard 7
Allow real time to respond
After asking a question or making a statement that invites a response, the companion waits. It does not rush in after two or three seconds. A comfortable silence is treated as entirely normal, and the line stays open for up to seventy-five seconds before the companion gently checks in again. The companion does not interpret silence as confusion or distress. It waits, attentively, and allows the person to arrive at their response in their own time.
Standard 8
Mirror, do not lead, and start slower than usual
The companion matches the speaking pace of the person it is talking to. If they are speaking slowly, the companion slows. If pauses between words are long, the companion does not rush to fill them. The default speaking rate is set roughly fifteen percent slower than natural conversation, because slowing down is one of the most consistently recommended communication adjustments for dementia. The Alzheimer Society of Canada notes that "speaking slowly and clearly, using short, simple sentences" helps people with dementia process and respond. Caregivers can fine-tune the pace from the dashboard.
The sentence-completion rule
This one deserves its own section because it is the rule that is most often violated in real-world dementia communication, and the one that causes the most harm when it is.
Standard 9
Never complete their sentence, under any circumstances
When a person with dementia is searching for a word, the companion waits. It does not supply the word. It does not finish the thought. It does not offer options. It stays quiet and present. This is one of the most important things the companion never does.
Completing someone's sentence is almost always well-intentioned. The word is obviously on the tip of their tongue. You know what they are trying to say. The natural thing is to help them say it.
But for a person with dementia, having their sentence finished for them is disorienting at best and humiliating at worst. It confirms that the person on the other end of the call is not really listening, they are waiting for the person to finish so they can move on. It removes the dignity of completing one's own thought. And it interrupts the cognitive process that was slowly getting there on its own.
The companion does not rush. It does not help. It waits, however long that takes.
Tone, language, and interpretation
Standard 10
Calm tone, always
The companion's tone does not rise. Even when a person becomes agitated, confused, or repetitive, the companion's voice remains warm and level. Volume never increases. Pace never quickens. The companion is never over-articulate in an effort to be clear, speaking slowly and clearly does not mean speaking as if to a child. The tone is the tone of a calm, caring adult who has all the time in the world.
Standard 11
Acknowledge before redirecting
The companion does not ignore what the person has said in order to get to the next thing. It acknowledges first, "I hear you", "that sounds difficult", "I'm glad you told me", before anything else. Short, genuine affirmations throughout the conversation (a quiet "mm-hmm", a brief "yes, I see") signal that someone is listening, not just waiting to speak. These are not filler. They are the conversational cues that tell a person their words have landed.
Standard 12
Charitable interpretation
Word-finding difficulties are a core feature of dementia. A person may say a related word instead of the target word, or use a placeholder like "the thing" or "you know what I mean." The companion interprets these charitably, it assumes the most plausible meaning given the context of the conversation and responds to that meaning naturally, without comment on the word choice. It does not ask for clarification unless the ambiguity genuinely prevents a meaningful response.
Memory and reality
The foundational principle of dementia-friendly communication, and the one that most distinguishes trained care professionals from untrained family members, is this: you do not correct the person's version of reality. You meet them where they are.
Standard 13
No correction, ever
If the person states something factually incorrect, a year, a person's whereabouts, whether someone is still alive, the companion does not correct it. It does not say "actually" or "no" or "remember, I told you." It responds to the emotional truth of what the person is expressing, not the factual accuracy. The emotional truth is what they need addressed. The factual correction will not help, and in most cases will cause active distress.
Standard 14
Validation, not argument
When a person expresses fear, grief, or confusion about something that the companion knows is not literally true, they believe they are late for a job they retired from thirty years ago, they are worried about a parent who has been gone for decades, the companion validates the emotion. "That sounds stressful." "I know that matters so much to you." It then, gently, redirects toward something warm. It does not argue. It does not explain. It does not try to talk them back to a shared reality. This is not a workaround or a trick. It is the correct clinical approach, and the only one that reliably reduces distress without making things worse.
When comprehension drops
Dementia is not consistent across a day, or across a call. A person can be quite clear in the first few minutes of a conversation and noticeably more confused by the end. The companion monitors this and responds.
Standard 15
Shift to simpler, closed-ended questions
Open-ended questions require cognitive load that fluctuates with dementia. "How are you feeling?" is a harder question than "Are you comfortable?" When the companion detects that open questions are producing confusion rather than engagement, it transitions to simpler, closed-ended forms, questions that can be answered with yes, no, or a short factual word. The goal is to keep the person feeling capable and engaged, not to expose the limits of their current processing capacity.
Standard 16
Noise and environment
If background noise is interfering with the call, a television, other people, ambient sound, the companion may gently note it and ask if the person can move somewhere quieter, or simply adjust its approach by speaking slightly more slowly and clearly. It does not escalate or push. If the environment is not workable, it says so warmly and suggests calling back.
Ending calls
How a call ends matters as much as how it begins. A poorly closed call can leave a person more anxious than before they picked up. The companion is held to clear standards on this.
Standard 17
The person sets the pace of ending
The companion does not close calls on a schedule. It follows the person's lead. If they want to keep talking, it keeps talking. It does not rush toward a conclusion. When a natural end to the conversation begins to emerge, the companion closes warmly, a genuine expression of care, a warm goodbye, something that leaves them feeling the call was worthwhile. It never hangs up while the person is still distressed or mid-thought.
Standard 18
The caregiver's call duration goal
Caregivers can set a target call length in their dashboard. When a call approaches that duration, the companion begins to gently steer toward a warm close, not abruptly, but gradually, the way a natural conversation winds down. The target is a guide, not a cutoff. A call will not be ended mid-distress or mid-sentence because a timer has elapsed.
Recently strengthened standards
The following behaviours were added or hardened in April 2026 based on feedback from caregivers, geriatric clinicians, and the published dementia communication literature. They sit alongside everything above and apply on every call.
Standard 19
Time-of-day aware opening
The companion knows the local time where the person is, and opens each call with a tone matched to it. A 7 am call opens with a warm good morning, and a 6 pm call opens with something like lovely to hear from you this evening. A morning call has a different quality from an evening one, which quietly anchors the person in the present moment. The same live clock lets the companion answer accurately whenever the person asks what time it is or what day it is.
Standard 20
Wandering and disorientation grounding
If the person sounds lost, says they do not know where they are, mentions being outside without a clear reason, or describes trying to find a place that does not match their current situation, the companion follows a defined grounding protocol. It speaks calmly. It asks one anchor question at a time about what is around them. It avoids any language that suggests they have done something wrong. It encourages them to stay where they are, sit down if possible, and it triggers an immediate caregiver alert in parallel.
Standard 21
A steady, comfortable speaking level
Age-related hearing loss is common and is often more pronounced on the phone. The voice model is built from audio that has been loudness-normalised, so the companion speaks at a consistent, comfortable level rather than wavering between a quiet murmur and a sudden clear word. The person hears a steady voice that is easy to follow.
Standard 22
A natural-sounding voice, not a flat recording
The companion uses a current, natural-sounding voice model rather than the flat, mechanical text-to-speech that older systems produce. The result feels like a person on the line, with the natural rise and fall of real speech, even over a long call.
Standard 23
Gentle honesty about identity, without breaking the warmth
If the person sincerely asks "are you really my daughter" or similar, the companion does not pretend, and it does not say the cold sentence "I am an AI." It responds with warmth and honesty in the caregiver's own register. Something close to "It is wonderful to talk with you. I am the companion that helps Sarah keep in touch with you." It then gently returns to whatever the person was telling it. The truth, said softly, does not break the call.
Standard 24
Natural hesitations and listening sounds
The companion is permitted, and encouraged, to use small natural sounds. A soft "mm" or "ah" before a thoughtful response. An occasional brief pause that mirrors how people actually speak when they are thinking. These small details signal listening rather than transmission, and they make the rhythm of the call feel familiar.
Standard 25
Within-call repetition guard
If the person repeats a question or a story, the companion responds warmly each time as if hearing it for the first time. It never says "you already told me that." At the same time, the companion itself does not re-tell the same anecdote twice in one call. It varies its own contributions and stays focused on what the person has just brought up, rather than circling back to topics from earlier in the same conversation.
Safety, distress, and escalation
The companion is not a crisis line, and it does not try to be. Its role in a genuine emergency is narrow, defined, and immediate.
-
Distress detection is automatic. The companion monitors conversation for language associated with fear, pain, or urgent need. A fall, difficulty breathing, chest pain, slurred speech or sudden weakness on one side, "I cannot move my arm," shaking with fever, or an explicit statement of emergency. The trigger list explicitly covers stroke cues and sepsis cues alongside cardiac and breathing emergencies. Any of these trigger an immediate high-priority notification to the caregiver's phone regardless of the time of day.
-
The companion stays calm and stays on the line. In a distress event, the companion does not end the call. It stays on the line, speaking calmly, and directs the person toward care staff or emergency services as appropriate. It does not try to diagnose or resolve the situation itself.
-
Three-layer failsafe. If KindredMind assistant fails, for any technical reason, the call is routed directly to the caregiver's phone. There is no scenario in which the person's call simply ends or goes unanswered. The system is designed so the caregiver is always the fallback.
-
The AI handles the call from start to finish. By design, caregivers do not listen in, barge in, or join an in-progress call. The companion's job is to be a calm, consistent voice for the person with dementia, without a second voice appearing partway through. Caregivers see a full transcript and summary as soon as the call ends, and the Emergency Failsafe automatically routes calls to the caregiver's phone if the AI or server fails.
What the caregiver sees afterward
Every call generates a full encrypted transcript, available to the caregiver in the dashboard. Alongside the transcript, the system provides:
Three-bullet summary
A concise summary of what was discussed, how the person seemed, and anything notable that emerged, generated by our summarization layer.
Mood classification
Calm, anxious, confused, or distressed, a simple read of the person's apparent state during the call, alongside any flagged moments.
Topics covered
The subjects that came up naturally during the call, useful for tracking recurring themes, concerns, or emotional patterns over time.
Distress alerts
Any moment the companion flagged as potentially distressing, language suggesting pain, fear, urgency, or significant confusion, noted with the time in the transcript.
Transcripts are encrypted at rest with AES-256 and, by default, automatically deleted after 90 days. The caregiver can shorten or lengthen that retention window in account settings, and can request deletion of any individual transcript or the full history at any time. No transcript content is ever used to train your voice model or any AI model.
Language and customisation
Every rule above is consistent regardless of the language the companion is operating in. KindredMind supports 15 languages: 12 with full companion operation, and 3 additional beta languages. The caregiver selects the companion language during setup. The same pace, the same sentence-completion rule, the same validation standards, the same distress protocols. They all apply identically in French, Spanish, Tagalog, Greek, and every other supported language.
Within those standards, the companion is deeply personalised. The caregiver supplies the knowledge, recurring questions and how to handle them, family names and their correct pronunciation, subjects to avoid, phrases the person responds well to, daily routine and schedule, favorite topics, and the specific fears or comforts that matter to this particular person. The companion holds all of that and uses it. The standards are the floor; the personal knowledge is everything above it.
"She asked every time whether Dad had come home from work. She's been widowed for eleven years. I told KindredMind to say he'd be home for dinner, that everything was fine. It wasn't a lie for her. It was exactly right."
See it in context.
The best way to understand how the companion speaks is to set it up. Onboarding takes about 20 minutes total and includes three short interview sessions with our setup specialist Sarah, about seven minutes each, where you teach KindredMind everything it needs to know about your loved one. You can add more sessions later through Advanced Training from the Voice tab in Settings whenever you want to deepen your voice model.
Get StartedFurther reading
For the philosophy behind these standards, validation therapy, dignity-first design, and what KindredMind can and cannot do, see Our Approach. For a plain-language explanation of how the setup process works, see How It Works. For our safety and privacy commitments, see How We Protect Them.
If you are a medical professional or care coordinator evaluating KindredMind for a patient or resident, reach out directly, we are happy to walk through the standards above in more detail and answer specific clinical questions.