Why Your Smart Speaker Is Eavesdropping On Client Calls

I’ve had versions of this conversation with enough people that it’s started to feel like a ritual. Someone mentions keeping a smart speaker in the home office — for background music, kitchen timers, quick weather checks — and almost without fail, another person brings up the theory: the device is listening all the time, feeding private conversations into some advertising machine. It sounds like paranoia, or at least it did until researchers started looking into what these devices actually do with the audio they collect. The truth, as it turns out, is more complicated than either the paranoid version or the reassurance you’d get from a product manual.

This post contains affiliate links.

The Eavesdropping Fear Has a Math Problem

Let’s start with the version of the fear that gets passed around most often — the idea that your smart speaker is constantly recording everything you say and feeding it into an advertising algorithm that serves you targeted ads based on your private conversations. Marketing firms have occasionally claimed exactly this. Two companies, CMG and Mindshift, reportedly bragged in client calls and podcasts about using smartphone and smart TV microphones to recognize keywords in conversations and build custom advertising audiences. CMG published a blog post stating that “Active Listening is often included” in terms of use agreements, though they later removed the post and apologized, claiming no eavesdropping was actually involved and that targeting data was “sourced by social media and other applications.” Mindshift similarly erased its marketing messages about this form of advertising.

But here’s where the math gets in the way. Google Cloud speech recognition pricing at its most discounted wholesale rate — two million or more minutes per month — costs about 0.3 cents per minute. If a device were processing, say, three hours of speech per day per user, the cost would land around $200 per year for that single person. Meanwhile, the annual advertising revenue platforms earn per user is less than $4 in some regions, around $10 globally on average, and up to roughly $60 in the U.S. The economics simply don’t support mass-scale audio surveillance for ad targeting. On top of that, modern operating systems show a clear indicator when the microphone is actively in use by an app, and continuous audio processing would drain battery and data in ways users would notice immediately.

Voice Data Collection Is Real — Just Not the Way You Imagined

Amazon states clearly that Alexa is always listening for its wake word, but only begins recording and sending voice data to the cloud after it hears that trigger, stopping when the interaction ends. That’s a different claim than “always recording,” but it still means snippets of audio from your home office — including conversations happening near the device — could end up on Amazon’s servers if something sounds close enough to “Alexa” or the wake word you’ve chosen. Independent studies confirm that voice data is used for ad targeting, and Amazon has not denied this practice.

The data collection doesn’t stop at ad targeting either. Apple Siri recordings were reviewed by human contractors for quality control and algorithm training, and Google Home Assistant recordings were similarly listened to by people. In 2023, the U.S. Federal Communications Commission secured a $30 million settlement against Amazon for failing to delete children’s data collected by Alexa, violating the Children’s Online Privacy Protection Act. The company was barred from using that illegally harvested data for business purposes, including training algorithms.

Researchers Christoph Lutz and Gemma Newlands, in a study of Amazon Echo and Google Home use in British households, found that concerns about third-party contractors listening to recordings were the most pronounced among users. Their work paints a picture of what they call “privacy pragmatism” or even “privacy cynicism” — users who are aware of the risks but continue using the devices anyway, often because they perceive the benefits as outweighing the potential harm. A separate systematic review of smart speaker privacy research by Guglielmo Maccario and Maurizio Naldi, published in the journal Security and Privacy, identified perception of privacy issues, vulnerabilities, and legal cases as core topics in a rapidly growing body of literature — confirming that these are not fringe concerns but central questions the research community is actively working through.

Shared Devices in a WFH Household Create a Different Risk

One in four U.S. adults and one in five Canadians own a smart speaker, according to survey data cited in a study from the University of British Columbia. That study, led by PhD student Yue Huang and senior author Konstantin Beznosov, looked at 26 Canadian adults who shared smart speakers at home — Amazon Echo, Google Home, Apple HomePod — and found that the nature of participants’ concerns depended heavily on their technical understanding of how the devices work.

People who were very familiar with the technology worried more about technical shortcomings: the smart speaker occasionally failing to distinguish one person’s voice from another, granting access to information it shouldn’t. Users with more basic knowledge focused instead on what housemates could potentially do — sometimes worrying about threats that didn’t actually exist on the device. One participant worried a family member could redial a phone number, unaware that the feature wasn’t available on their particular model.

For anyone working from home in a household where other people — partners, older children, housemates — have access to the same smart speaker, this creates a concrete risk that has nothing to do with corporate surveillance. A housemate could unintentionally order something using your account, overhear a private work conversation, or access reminders and calendar entries that contain client information. A child could accidentally use a last-number-dialed feature to call a client or employer. These are not hypothetical edge cases; they are the kind of everyday accidents that happen in shared living spaces.

Client Calls and the Question of Third-Party Access

Here’s where the WFH-specific angle sharpens. A conference paper presented in 2021 by researchers Andreas Liesenfeld, Gábor Parti, and Chu-Ren Huang demonstrated that conversational AI can predict speaker traits after eavesdropping for only 30 seconds. The study focused on what can be inferred from voice patterns — not content, but characteristics like personality, emotional state, or demographic markers. In a separate stream of research, voice biomarkers have shown promise in diagnosing heart failure and coronary artery disease from subtle speech parameter changes. The point is not that your smart speaker is diagnosing your clients, but that voice data carries far more information than the words themselves.

If a snippet of a client call is accidentally captured because someone in the next room said something that triggered the wake word, that recording — or a transcript of it — could end up stored in the cloud, reviewed by a contractor, or used for model training. The Lutz and Newlands study found that concerns about third-party access were the most intense among users, and for good reason: the chain of custody for voice data is longer than most people realize. It doesn’t just sit on your device or even on the vendor’s server. It may be accessed by moderation teams, used to improve speech recognition, or analyzed for ad targeting.

For professionals who handle confidential client information — lawyers, therapists, accountants, consultants — the implications are straightforward. A voice assistant in the same room as a client call creates a data flow that you didn’t authorize and may not be able to control. No amount of convenience from voice-controlled calendar checks or hands-free note-taking justifies exposing privileged communications to a system whose data handling you can’t audit.

Practical Steps That Don’t Require Throwing Out Your Speaker

You don’t necessarily have to banish every smart speaker from your home to protect client calls. But you do need to be deliberate about where and when these devices are active. Here’s what a realistic approach looks like.

If you share your home office space with other people, the conversation needs to include them too. Saying “hey, could you avoid using the smart speaker while I’m on client calls?” is a start, but it helps to be specific about why. Try something like: “I handle confidential information on these calls, and if the speaker picks up even a few seconds of audio, that data could end up stored on a server I can’t control. I need us to unplug it during my working hours.” That gives the other person a reason rather than just a rule.

For those who want to keep the convenience of a smart speaker without the privacy trade-off, the honest answer is that you’re accepting some level of risk. The question is whether that risk is proportionate to the benefit. A systematic literature review on smart speaker privacy noted that perceived usefulness and enjoyment are strong drivers of adoption — but also that perceived privacy risks are a significant negative factor, particularly in countries like the U.S. where awareness is higher. You get to decide where you land on that balance. The important thing is to land there intentionally, not because you assumed the device was harmless or because you assumed it was spying on you in ways it isn’t.

The most practical takeaway from all of this research is also the least satisfying: the risk is real but it doesn’t look like the conspiracy theory. It looks like a device that occasionally mistakes a conversation for a command. It looks like a shared household where someone else accesses your calendar. It looks like a contractor in another country listening to a few seconds of a client call because the system flagged it for quality review. None of those scenarios require a corporation to be secretly recording everything you say. They just require the device to work the way it was designed — and for that design to not fully account for the privacy needs of someone who takes confidential calls from their living room.

You can work around it. Unplug the speaker during calls. Turn off the microphone on your laptop when you’re not actively using it. Talk to the people you live with about what’s at stake. None of this is complicated. It just requires treating the smart speaker like what it actually is: a convenience that comes with a trade-off, not a threat and not a toy, but a piece of hardware that doesn’t know the difference between a grocery list and a privileged conversation between you and your client.

That distinction is yours to manage. The research just tells you where to look.