0 / 6

You can call Aika on WhatsApp now

Aika now answers WhatsApp calls and calls people back on WhatsApp. Why we built it, what changes in the audio, and the two bugs that only a real call could show us.

Contents

Open the chat with Aika on WhatsApp and tap the call button. Aika picks up, the same Aika you get on a phone call, and when WhatsApp shares your number with us, the same memory of you. If you called on WhatsApp, Aika calls you back on WhatsApp too, when WhatsApp allows it: reminders, callbacks, the call after a dropped line. If you miss that WhatsApp call, a normal phone call follows about a minute later. If you decline it, nothing follows.

That is the whole feature from the caller's side. This post is about why we wanted it, what it changes in the audio, and what it took to make it work on real calls.

Why WhatsApp

A lot of people already message Aika on WhatsApp. For many of them a WhatsApp call is the natural way to talk to someone: it costs nothing on Wi-Fi, it works when the mobile signal is weak but the data isn't, and it is where their family already calls them.

There is also a reason that matters to us more than it should matter to them. In an earlier post we explained why we could not get wideband audio over Indian phone lines: every hop of the call has to carry it, and most don't. A WhatsApp call has no such chain. The audio goes from the app to us as Opus over the internet, and nothing in between squeezes it back to 8 kHz.

What changes in the audio

A teammate called Aika once on WhatsApp and once on an ordinary phone call, and we compared what our speech recognition actually received on each, from the same speaker.

BandWhatsApp callPhone call
300 to 3,400 Hzmost of the energymost of the energy
3,400 to 4,000 Hzreal speech, about 28 dB below the totalclose to nothing, about 43 dB below the total
above 4,000 Hzspeech rolling off smoothlymostly conversion artefacts

The phone line does what phone lines do: it cuts off a little above 3.4 kHz, which is where much of the difference between "s", "sh" and "f" lives. The WhatsApp call keeps that band. We currently cap our WhatsApp audio at 16 kHz sampling, so it carries sound up to about 7 to 8 kHz. That is roughly the "HD voice" range a VoLTE call can reach when every hop cooperates, without needing every hop to cooperate.

We are careful about what this proves. In the wideband post, cleaner audio did not by itself make our recognizer make fewer mistakes on the calls we could test. It is the same here: the audio is clearly better, and whether that turns into fewer misheard words is something we will measure on real calls rather than assume.

Two ways in, and the one that worked

WhatsApp's business calling can reach you in two ways. One hands you the call as SIP, the protocol phone networks use. The other sends you a webhook when a call arrives, you answer through an API, and the audio runs over WebRTC, the same technology browsers use for video calls.

We built the SIP version first, because our call gateway already speaks SIP to the phone network. It worked on the first test call, and then most calls after it failed: the call would connect, and the audio would never start. We spent a morning ruling things out (our network, our certificates, our audio format, our code, even a second independent SIP client we wrote just to check) and the failure followed every one of them. Then we tried the API route with a small throwaway script, and the first call had audio both ways. We moved our gateway to the API route the same day.

The bugs only a real call shows

Both of these passed every test we had, and both broke on the first real call.

A codec setting nobody reads. Our gateway told WhatsApp it would accept Opus with a minimum packet time of 10 milliseconds. WhatsApp's calls ask for 20. To the WebRTC library we use, that difference made them two different codecs, so it refused to start the audio at all. Every call failed within a second or two. Our tests passed because our fake WhatsApp had been written with our own setting, not WhatsApp's. The fix was one line. The lasting fix was making the fake copy what the real thing sends, so the test fails the way the real call did.

An answer before anyone answers. When Aika calls someone on WhatsApp, WhatsApp sends us the call's audio setup while the person's phone is still ringing, and only a few seconds later tells us they picked up. We had treated the first message as "answered", so Aika started speaking into a ringing phone, and a missed call looked like a finished one, which meant the phone call that should follow a missed WhatsApp call never came. Now Aika waits for the pick-up before it speaks, and a missed or declined call is reported as exactly that.

Neither bug was exotic. Both were cases where we had written down what we expected the other side to do, and the other side did something slightly different. A test that uses your own guess about the other side cannot catch that, so for this kind of integration a real call stays part of every release.

What's next

WhatsApp calling is on for everyone who calls Aika's WhatsApp number, and callbacks go out on WhatsApp to everyone who last called that way, whenever WhatsApp allows the call. Next we want to see, on real calls, whether the wider audio means fewer misheard words, especially for names, and for the sounds a phone line blurs together.

References

Cite as

Aika Labs. 'You can call Aika on WhatsApp now.' aikalabs.ai/blog/calling-aika-on-whatsapp, 2026-11-22.

Related