← back to blog

Speech Synthesis Voices and What They Reveal

Most people who run multiple browser profiles think about canvas, WebGL, and fonts first. Speech synthesis voices barely come up, and that’s exactly why they’re worth a look. It’s a small API, it’s easy to query, and the data it hands back is more specific to your real machine than most people expect.

What the speech synthesis API actually exposes

Every modern browser ships a window.speechSynthesis object as part of the Web Speech API. It exists so pages can trigger text-to-speech, for accessibility tools, read-aloud features, language learners, that kind of thing. Call speechSynthesis.getVoices() from any page’s JavaScript console and you get back an array of voice objects. Each one has a name, a language tag, a voiceURI, and a couple of booleans for whether it’s a local voice or requires the network, and whether it’s the default.

None of that requires a permission prompt. A site doesn’t ask before reading your voice list, the same way it doesn’t ask before reading your screen resolution. It’s just there, sitting in the DOM, waiting for a script to call it.

The interesting part is where those voices come from. They’re not bundled with the browser. They come from the operating system’s text-to-speech engine, plus whatever the browser itself adds on top. On Windows that means the SAPI5 voices tied to your installed language packs, names like Microsoft David or Microsoft Zira, plus whichever additional voices got installed with Windows updates or language settings. On macOS it’s the system voices from the Speech framework, dozens of them if you’ve never touched the settings, a much shorter list if you have. Chrome and Edge also inject their own network-based voices on top of the OS list, which is a separate signal from the local ones.

So a page that reads your voice list is really reading a cross-section of your OS version, your installed language packs, your regional settings, and which browser you’re on. That’s a lot of information from one API call that most users never notice exists.

Why the voice list makes such a clean fingerprint signal

A few properties make this list unusually useful for tracking, compared to other fingerprint surfaces.

First, it’s stable. Your canvas fingerprint can shift a little with GPU driver updates. Your voice list doesn’t change unless you actually install a language pack or upgrade the OS. That stability is valuable to a fingerprinting script, because a signal that stays the same across sessions is a signal that can tie two visits together with confidence.

Second, the count and exact names are specific. A stock Windows 10 install in English might expose two voices. A machine where someone has installed several language packs for work could expose fifteen or twenty. That range, plus the exact names and order they come back in, narrows down the pool of machines that could produce that exact list. It’s not unique on its own the way a login cookie is, but combined with two or three other signals it tightens things up fast.

Third, it’s cheap to check and hard to fake convincingly, because it has to line up with everything else the browser is reporting. If a page sees a user agent claiming macOS but a voice list full of Microsoft SAPI names, that’s not a subtle mismatch. It’s a direct contradiction that takes one line of code to catch.

Where it fits in the bigger fingerprint picture

Speech synthesis voices are rarely the whole story. Detection systems build a composite picture: canvas rendering, WebGL renderer strings, installed fonts, audio context output, screen and viewport dimensions, timezone versus IP geolocation, and the voice list all get folded together into one profile. Any single mismatch might get shrugged off. Several mismatches pointing the same direction, OS claims one thing, fonts suggest another, voices suggest a third, and a script has enough to flag the session as inconsistent, even before it looks at behavior.

This is the part that trips people up when they’re setting up profiles for multi-accounting. The instinct is to focus on the signals that are easy to see, like the user agent string or the IP address, and leave the quieter APIs alone because nobody thinks to check them. Voice synthesis is one of the quieter ones. It sits there exposing OS-level detail while the operator is busy tuning the more visible stuff.

Where anti-detect browsers try to intervene

Tools built for running many isolated profiles, Multilogin, GoLogin, Kameleo, AdsPower, Dolphin Anty, and others in that category, generally work by building a browser profile around a target fingerprint: a specific OS, browser version, and hardware combination, then adjusting the values that JavaScript can read to match that target. Canvas and WebGL noise injection are the parts people talk about most, but a profile that’s actually trying to be internally consistent also needs to handle the voice list, because otherwise it’s the one API left unmasked while everything else got dressed up.

The sound approach is to make the exposed voice list match the OS the profile claims to be. A profile presenting itself as Windows should return Windows-style SAPI voice names in a plausible count, not the macOS system voices left over from the machine the software is actually running on. Some tools handle this by overriding getVoices() at the API level and returning a canned list appropriate to the spoofed OS. How well that’s implemented varies by product and by how often the underlying voice data gets updated to track real-world OS releases, and it’s the kind of detail that’s easy to get half right, matching the OS but not the browser version, or getting the count right but not the exact naming convention a real install would produce.

None of that makes a profile undetectable. It makes one narrow signal harder to use in isolation. A detection system that’s already suspicious for other reasons, unusual timing patterns, a login history that doesn’t match the claimed device, will keep looking regardless of whether the voice list checks out.

What to actually check when you’re testing a profile

If you’re evaluating whether a given anti-detect setup handles this properly, the check itself is simple and doesn’t require bypassing anything, it’s just reading values the page already exposes to you. Open the profile’s browser, hit the console, and run speechSynthesis.getVoices(). Compare what comes back against the OS and browser the profile claims to be. Does the voice count look like a realistic Windows or macOS install rather than an empty array or an obviously wrong list carried over from the host machine? Do the voice names match the claimed platform’s naming conventions? Does it stay consistent if you reload the page, since some implementations only populate the list correctly after the async voiceschanged event fires, and a broken override can leave the array empty on first call?

It’s also worth checking this alongside the other signals rather than in isolation. A voice list that matches the claimed OS but sits next to a WebGL renderer string that clearly doesn’t is still a contradiction, just spread across two APIs instead of one. Internal consistency across the whole fingerprint is the thing that matters, not any single value in isolation.

The honest takeaway

Speech synthesis voices are a small piece of a much larger picture, and treating any single fix as the thing that makes a profile safe misses how detection actually works. No configuration, override, or piece of software makes a browser profile undetectable or a multi-accounting setup risk-free, and platforms update their detection logic regularly enough that anything held up today as a definitive fix can stop working without warning. What’s worth doing is understanding which signals exist, checking them yourself instead of taking a vendor’s word for it, and treating consistency across the whole fingerprint as the goal rather than chasing any one API.

If you want more breakdowns like this on the signals real detection systems actually check, and hands-on looks at how different anti-detect browsers handle them, head back to the Anti-Detect Review home page.

Get new guides and videos first — join the Telegram channel.

need infra for this today?