AudioContext sampleRate and latency as a fingerprint

An AudioContext fingerprints a machine before it renders a single sample: sampleRate, outputLatency and destination.maxChannelCount are read straight from the operating system’s audio driver, so a headless server with no sound hardware answers them with tell-tale defaults. The defence is not to spoof each number, but to return all three as one coherent device profile that a real machine could actually have.

Most writing about audio fingerprinting is about the rendered buffer: you run an OfflineAudioContext, hash the samples it produces, and treat that hash as the signal. That is a real surface and it has its own page here.

This page is about the other half, the part that needs no rendering at all. Before a single sample is generated, an AudioContext already exposes three plain numbers that come straight from the operating system’s audio stack: sampleRate, outputLatency, and destination.maxChannelCount. They are cheaper to read than any buffer hash, they are stable within a session, and on a headless server they are one of the clearest tells that no real sound hardware is present.

The three numbers a real audio stack exposes

Open an AudioContext on a normal desktop and read four fields:

from invisible_playwright import InvisiblePlaywright

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com")
    profile = page.evaluate("""() => {
        const ctx = new AudioContext();
        const out = {
            sampleRate:      ctx.sampleRate,
            outputLatency:   ctx.outputLatency,
            baseLatency:     ctx.baseLatency,
            maxChannelCount: ctx.destination.maxChannelCount,
        };
        ctx.close();
        return out;
    }""")
    print(profile)

Each of those fields is a question to the driver, not to the browser:

• sampleRate is the rate the audio device is running at, almost always 44100 or 48000 Hz. It is decided by the sound card and its driver, not by the page.
• outputLatency is roughly how long it takes a sample to travel from the graph to the speaker. It reflects the buffer size the driver negotiated, and it differs between onboard audio, a USB interface, and a surround setup.
• maxChannelCount on the destination node is how many output channels the device can drive: 2 for ordinary stereo, 6 for a 5.1 setup, 8 for 7.1.

None of these is an automation flag. They are hardware facts, and that is exactly what makes them useful to a detector: JavaScript cannot invent a sound card it does not have.

Why the values have to agree with each other

The trap here is not any single value being unusual. Every one of these numbers, taken alone, is common. The signal is in the combination.

Real devices come in a small number of shapes. Onboard laptop audio tends to run stereo at 48000 Hz with a modest latency. A cheap USB headset might report 44100 Hz. A discrete sound card feeding a 5.1 receiver reports six channels and a latency in a range that matches a card built for that job. What you never see on real hardware is the cross-product: a maxChannelCount of 6 paired with the sample rate and latency of a thin laptop codec, because no such device exists.

That cross-value contradiction is precisely what a consistency-minded detector looks for, and it is the same class of check that catches a Windows user agent over a Linux font set, or hardware concurrency and device memory that pair oddly with the machine being claimed. A fingerprint is not a list of values, it is a set of relationships, and the relationships are where a naive spoof falls apart.

What a headless container reports instead

Run the same three reads inside a server container with no sound hardware and you get one of two failure shapes, both of them loud.

The first is a null audio backend: the context comes back with a default sample rate and a latency that is suspiciously round or zero, the same on every container image, because there is no driver underneath negotiating anything. A value that is byte-identical across thousands of visitors is not neutral, it is a cohort.

The second is worse. Some headless setups fail to construct the context at all, or suspend it immediately, so the page sees a state that never reaches running. As with every other surface, an empty or blocked result is itself a signal, not a clean pass. A detector that records suppression by name counts the absence against you exactly as it would count a wrong value.

This is the same reason an empty speech-voice list gives a server away: both audio output and installed voices are reads into the operating system, and a datacenter answers them with defaults that say “datacenter”.

How this browser keeps the profile coherent

The fix is not to pick three plausible numbers independently, because independent draws are what produce the impossible device. The engine returns all three together as one coherent profile, chosen for the session from a small set of real device archetypes, so the values always describe a machine that could actually exist:

Device archetype	sampleRate	outputLatency	maxChannelCount
Plain stereo desktop	44100 Hz	around 40 ms	2
USB audio interface	48000 Hz	lower, near 30 ms	2
5.1-capable sound card	48000 Hz	near 40 ms	6

Each row is a machine that exists; the impossible combinations between rows are the ones a detector is looking for.

Because the whole profile is keyed off the session seed, a six-channel destination never lands next to a laptop’s latency, and the same seed reproduces the same audio device on every run. The three fields move as a unit, which is the property the detector is testing for. This is the general principle these patches follow: the goal is to look like a real machine, not merely to avoid an obvious default, and a suppressed or arbitrary value fails that goal as surely as a wrong one.

Measuring it yourself with a fixed seed

The point of a seed is that a finding is reproducible. Read the profile under two seeds and confirm two things at once: that each profile is internally coherent, and that the same seed returns the same device every time.

from invisible_playwright import InvisiblePlaywright

READ = """() => {
    const ctx = new AudioContext();
    const out = [ctx.sampleRate, ctx.destination.maxChannelCount,
                 Math.round((ctx.outputLatency || 0) * 1000)];
    ctx.close();
    return out;
}"""

for seed in (42, 1337):
    with InvisiblePlaywright(seed=seed) as browser:
        page = browser.new_page()
        page.goto("https://example.com")
        rate, channels, latency_ms = page.evaluate(READ)
        print(f"seed={seed}: {rate} Hz, {channels} ch, ~{latency_ms} ms")

Two checks turn this into evidence rather than a single reading:

1. Run each seed twice. The tuple must be identical across runs of the same seed. If it drifts, something is randomising per call, which is the cheapest tampering check a detector has.
2. Sanity-check the pairing by hand. Six channels should never appear beside a latency that belongs to onboard stereo. If it does, you have found the exact contradiction this section exists to prevent.

Compare the output to a stock Firefox on the same machine and you will see the shape of the guarantee: the stock browser reports your real card, the seeded browser reports a coherent invented one, and neither reports the round, hardware-free defaults a bare container does. If you need a specific device rather than a seed-derived one, pinning a field leaves the rest of the identity untouched.

Conclusion

sampleRate, outputLatency and maxChannelCount are read straight from the OS audio driver, which means a headless server has nothing truthful to say through them and a careless spoof says something impossible. The buffer hash gets all the attention, but these three static numbers are read first and cost nothing to check. Treating them as one coherent device profile, reproducible from a seed, is what keeps a six-channel count from ever sitting next to a laptop’s latency, and that pairing is the thing a consistency check is actually built to find.

Short answers to the questions that lead here

What is AudioContext.sampleRate and can I change it from JavaScript? It is the rate the audio device runs at, 44100 or 48000 Hz, read from the driver. Page JavaScript cannot change it truthfully; only the engine underneath can decide what the read returns.

Why does my headless server report the same audio values on every run? Because there is no sound hardware, so the context falls back to a fixed default. Identical values across every container are a cohort, not a clean fingerprint.

Is outputLatency a reliable fingerprinting signal? On its own, weakly. In combination with sample rate and channel count it is strong, because only certain triples correspond to devices that actually exist.

What does maxChannelCount reveal? How many output channels the device can drive: 2 for stereo, 6 for 5.1, 8 for 7.1. A value that contradicts the latency and rate is the tell.

How do I make these values consistent with each other? Do not draw them independently. Pick a real device archetype and take all three from it, which is what a seed-derived profile does automatically.

Will an empty or suspended AudioContext pass a detector? No. A blocked or suppressed read is recorded as its own signal, so an absent value fails the same way a wrong one does.

Sources

• W3C Web Audio API specification, which defines sampleRate, baseLatency and outputLatency on AudioContext and maxChannelCount on the destination node: www.w3.org/TR/webaudio/.
• MDN Web Docs on the same properties, read from the engine rather than from any single detector’s rendering of them: AudioContext.sampleRate, AudioContext.outputLatency, and AudioDestinationNode.maxChannelCount.
• This project’s own audio patches and their release notes, which choose the three static audio properties together as one coherent device profile so the values never contradict each other.

See also: AudioContext fingerprinting for the rendered-buffer side of the same surface, why getVoices() comes back empty for the other operating-system read that gives a server away, and the checklist for being detected on one site for where the audio device sits in the order you should debug.

---

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. The audio device is one of the machine tells no page-level plugin can reach, which is why it is solved in the engine.