Slapback · 2026-09-20

Double-talk: why echo cancellers clip your words

By the Echobreak team

You know the moment. Someone's wrapping up a point, you jump in with "actually—" and the call delivers "—tually." Your first syllable didn't get lost in the network; it was eaten, deliberately, by the same software that keeps your calls echo-free. The situation even has a name in the engineering literature — double-talk — and it's widely considered the hardest thing an echo canceller does. Here's what's happening in that half-second, and why speakerphone setups feel it most.

The canceller's impossible moment

Remember the canceller's job: the mic signal contains echo (the far voice, out of your speakers, back into your mic) and it must subtract that echo using its learned model of your room. While only the far side talks, this is comfortable — everything at the mic is echo, subtract with confidence, transmit silence.

Now you interrupt. For the first time, the mic carries two things at once: the echo of their voice and your genuine speech, overlapping in time and frequency. The canceller must remove one completely while leaving the other untouched — separating two voices sharing the same air, in a few milliseconds, without tomorrow's hindsight. That's double-talk. Three failure modes are on the table:

Under-cancel: protect your speech, let echo residue through — the far side briefly hears themselves. The cardinal sin; cancellers are tuned hard against it.
Over-suppress: when unsure, duck everything. Safe from echo — but your syllables die with it. This is your missing "ac-".
Corrupt the model: the sneakiest one. The canceller learns your room by watching mic-vs-played correlations; if it keeps "learning" while you're talking, your voice poisons its model and echo leaks for the next few seconds. So real cancellers run a double-talk detector whose job is to freeze learning the instant near speech appears — and every detector has reaction time.

Add these up and the everyday symptoms explain themselves: first syllables clipped on interruptions (the detector needed a beat to decide you were real), voices ducking in and out during crosstalk (over-suppression winning), a breath of echo right after a heated exchange (the model took a hit and is re-learning).

Why muted-mic calls never feel it

Double-talk needs a mic that's live while the far side plays through speakers. Headphone users barely trigger it (no speaker signal at the mic — the canceller has nothing to guard against), and everyone-muted meetings never do (no near speech). This is why the clipping complaint clusters so heavily among speakers-only people: they're running the canceller in its hardest mode all day. It's also the honest asterisk on our own no-muting advocacy — an open-mic, open-speaker call is asking the software to do its most difficult trick continuously, and the polish of that trick is a real quality difference between cancellers.

What actually helps

Lower speaker volume. Everything about double-talk is a ratio fight between your voice and speaker bleed at the mic. Less bleed = easier separation = softer hand on the suppressor. Same first move as every echo problem, for the same underlying reason.
Get the mic closer to your mouth. Raises your side of the ratio. A close mic at low gain is the single best anti-clipping configuration open-speaker life allows.
Interrupt like radio operators do. Half-serious, fully effective: lead with a throwaway sound — "yeah—so," "right—but" — and let the detector spend its reaction time on the throwaway. Your content survives.
Try a different canceller. Double-talk handling is where implementations differ most — it's compute-hungry, so built-in cancellers economize on it. If clipping is chronic on your setup, a dedicated layer with modern separation (the tools in our guide) is the variable worth testing; free tiers make the test cost nothing.
Accept the physics gracefully. Even perfect software can't transmit a syllable it hasn't decided is speech yet. If a meeting is genuinely interruption-heavy and matters, headphones for that hour remove the problem class entirely — no shame in it.

Next time a call eats your first word, you'll know: somewhere in your machine, a detector spent forty milliseconds deciding whether you were a person or a reflection. It guessed conservative. It usually does.