0 / 5

Why wideband phone calls are out of reach in India, for now

Wideband audio should help speech recognition on phone calls, so we tried to get it for our callers. Every hop of an Indian call has to agree, and most don't.

Contents

A phone call in India usually reaches us as 8 kHz audio. That keeps sound up to about 3.4 kHz and drops most of what is above it, and above it is where a lot of the difference between "s", "sh" and "f" lives. If you are building speech recognition for people calling from phones, the obvious wish is wideband: 16 kHz audio, sound up to about 7 kHz, the "HD voice" a lot of phones already use.

Earlier this month we spent a few days trying to get it for Aika's callers. We were fairly sure it would help. This is what we found, and why we stopped.

Every hop has to carry it

In the previous post we followed a call from the handset to our speech model. Wideband only reaches you if every link on that path carries it. A single narrowband hop anywhere and you get narrowband, however good the rest of the path is.

PhoneVoLTE HD: wideband. 2G or feature phone: narrowbandBetween operatorsusually narrowbandTrunk to youG.711 by default; G.722 only if the operator delivers itAll three must agree. Open marks are where it usually fails.PhoneVoLTE HD: wideband. 2G or feature phone: narrowbandBetween operatorsusually narrowbandTrunk to youG.711 by default; G.722 only if the operator delivers itAll three must agree.
Wideband reaches you only if every hop carries it. One narrowband hop anywhere and the call arrives as 8 kHz audio.

The caller's phone. A 4G smartphone on VoLTE with HD voice on can send AMR wideband. A 2G call in India in practice can't, and neither can most feature phones. A lot of the people we most want to reach call from exactly those phones.

Between operators. When the caller is on one network and your number is on another, the call crosses an interconnect, and those crossings are commonly narrowband. So even a caller whose phone can do wideband usually gets squeezed back to 8 kHz there.

The trunk to you. Your number sits on a SIP trunk. For wideband to reach your code, the operator has to hand the call over as a wideband codec, usually G.722. The default handoff in India is G.711.

What the trunks actually offer

We went through the published technical documentation of every provider we could find that sells SIP trunks with Indian numbers. Most list only G.711 (PCMA and PCMU). One operator's FAQ mentions G.722 while describing bandwidth needs. One international provider lists G.722 among its trunk codecs in general, but says nothing about whether it holds on calls to and from the Indian phone network. A few pages mention G.722 in a general explainer with the caveat "if the carrier supports it", and then offer no setting anywhere to turn it on.

So the realistic best case looks something like this. Put a number on a trunk with one operator, get that operator to confirm in writing that calls arrive as G.722, and then a caller on that same operator, on VoLTE, with HD voice working, might reach you in wideband. Everyone on another network, and everyone on 2G, still arrives narrowband. Covering them would mean a separate number and trunk on every operator, plus some way to steer each caller to the right one. You don't control that. The caller dials a number and their network decides the route.

So we measured

Documentation only gets you so far, so we set up a test trunk, called it from our own phones and captured the raw packets.

The test trunk only offered G.711. The call negotiated A-law, with zero packet loss and a clean signal. No wideband.

Then we compared it with the same person, phone and sentence on our existing route, and two things stood out. The existing route had a faint, regular crackle that came from a conversion step and not from the network (more on that in the previous post). It also had noticeably more energy above 4 kHz, which sounds like a win until you remember that narrowband audio should have almost nothing up there. That energy was junk.

What we actually cared about was whether either route made fewer recognition errors, so we ran the same recordings through our speech recognizer two or three times each, on each route, both raw and filtered.

No route won. Every version misheard something, and the mistakes repeated on every run of the same recording, so it was not random noise from the recognizer. They moved between a person's own calls much more than they moved between routes. One call misheard a proper name on one route, and another call from the same person on the same phone misheard an ordinary phrase on the other. Filtering out the junk above 4 kHz fixed neither.

That result is what stopped us. The cleaner route gave us cleaner recordings, which is useful for reviewing calls and for training data. It did not give us fewer errors, at least not on the calls we could test, and switching would have cost real effort for a gain we couldn't show.

What we do instead

Narrowband isn't going away for our callers any time soon, so we treat it as the condition we design for.

We evaluate on phone audio. A recognizer that scores well on clean 16 kHz recordings tells you very little about an 8 kHz call from a busy street, so every speech recognition change we make gets tested on real narrowband calls and graded by people who speak the language.

We stop adding damage where we can. We can't make the network wideband, but where the 8 to 16 kHz step is ours, we do it with a resampler that remembers the previous packet.

We give the recognizer context. A lot of errors on a phone line are one word that sounds like another, and knowing what was just asked helps more than a few extra kilohertz would on most calls.

And we look at errors per call instead of per vendor. Ours varied between calls more than between providers, which points at the caller's network, room and phone. Those change every time, and no trunk contract changes them.

Wideband will come as VoLTE spreads and interconnects improve. When an operator can tell us in writing that its trunk delivers G.722 on real calls, we will test it the same way: same phones, same sentences, recognition errors counted. Until then the work is getting good at 8 kHz.

References and further reading

Cite as

Aika Labs. 'Why wideband phone calls are out of reach in India, for now.' aikalabs.ai/blog/why-wideband-calls-are-out-of-reach-in-india, 2026-11-14.

Related