Improving Audio Quality in Duo with WaveNetEQ | Hacker News Reader

Improving Audio Quality in Duo with WaveNetEQ

ai.googleblog.com

93 pointstheafh35 comments

I can't wait for gradual improvements in deep learning allow video calls to extrapolate my presence for longer and longer time segments, until my connection can drop out for minutes and WebRTC will make it convincingly seem like I'm still participating in the meeting, making reasonable suggestions for possible code architectures and screen sharing hallucinated PowerPoint presentations.

More seriously I worry about cases where it hallucinates the wrong part of a word and changes the meaning of what I was trying to say. Although we kind of already live in this world with phone autocorrect and it's mostly fine. I'm weird though and disable autocorrect on my phone so that when I make a typo it's obviously a typo instead of just saying a plausible but wrong thing.

> More seriously I worry about cases where it hallucinates the wrong part of a word and changes the meaning of what I was trying to say.

Xerox had this issue in 2013

https://www.theverge.com/2013/8/6/4594482/xerox-copiers-rand...

https://news.ycombinator.com/item?id=6156238

It was easy to tell when autocorrect switched to machine learning approaches: instead of getting the basics correct and occasionally missing long words or proper names, it immediately and aggressively began to attack correct sentences, committing grade-school grammar mistakes on your behalf if you weren't vigilant enough to catch it in the act and revert its "helpful" changes.

Or photocopiers changing numbers on you because the compression algo thought that 6 was an 8:

https://www.bbc.co.uk/news/technology-23588202

I tried a couple of times to use Dragon Naturally Speaking to author reports for work. I quickly realized it never made a spelling error, but that definitely didn’t mean that it was 100% correct. I ended up having to read backwards through the report in order to pick up mod most of the errors.

Ultimately it was way too much work and too high risk for customer facing content, so I would use it every now and then help me unblock my thinking for a particular topic, but I had to resort to typing for the real deal.

Why the duck would you do that?

Duo is sad, and I feel Google is missing the point with all their tech improvements with Duo but no usability improvements.

I bought a Nest Hub Max for kids and grandparentst for video calls using Duo with family, and for one on one calls it's great especially for kids who move around, but with quantine the whole world (my version!) has quickly moved on to group video calls which ppl can enter and leave as they please.

Consequently everyone has stopped using Duo as it doesn't support that, and turns out you need to be able to easily chat with your calling group, another thing duo doesn't support.

Zoom, WhatsApp, Facebook Messenger all work much better from a talking to a family group with varying tech know-how.

Zoom is a dedicated video but works much better than duo as you can make a link and drop it in a chat group for ppl to join. Duo has no such features!

So my fancy duo called machine is now just a photo display.

Worse, Hangouts supported that at the time Duo launched. Duo was a downgrade.

Duo supports group calls on mobile devices and browsers FYI.

But you can't add or remove people in a call. You have to call everyone simultaneously up front. If you didn't add someone at the start, you have to hang up and start a new call, and hope that everyone else hangs up too so they can join the new call.

Duo has a ton of problems around establishing calls, and identity in general especially with multiple devices. It also has kind of a mystery meat UI with unlabeled buttons and undiscoverable interactions. And the artificial separation between text and video chat (Allo/Duo) never made sense in the first place, and continues to hamstring Duo even after Allo's death.

Not on browsers, mobile, tablet and ChromeOS running the android app.

I use Signal and Duo a lot to talk to my Dad in India and I can say even with low bandwidth when nothing else works, Duo audio calls are crisp clear.

I work on calling at Signal. If Signal worked for calling your dad in India, would use it instead of Duo? If so, would you mind helping us improve it by providing feedback for new builds and that sort of thing? If so, email me at peter@signal.org. Detailed feedback from you might help a lot.

Signal used to not care much about call quality, I'm glad to see a massive cash infusion is turning that around!

Sent my feedback to your email.

They lead with some very interesting numbers:

" 99% of Google Duo calls need to deal with packet losses, excessive jitter or network delays. Of those calls, 20% lose more than 3% of the total audio duration due to network issues, and 10% of calls lose more than 8%."

This sounds very puzzling. If those are packet loss numbers, those conditions shouldn't be able to carry normal TCP (eg HTTP/HTTPS) traffic properly. And these are the people ambitious enough to try Duo so probably 20% would be a lower bound for the unusably-bad-network user share.

Is there published research to correlate these numbers with?

Can there be other explanations? Eg people on mobile connections that just cut out for long periods of time in middle of calls?

TCP is self correcting. So in the end the receiver gets the whole data, correctly. But it happens with a delay.

For audio you need real-time synchronous data feed, and cannot wait for the error correction round trip to happen.

IMO you absolutely can wait for round-trips, etc. The problem arises when delay is not consistent. But if the delay is consistent, it's not a big problem. When I was child and called my relatives in another country, that's how it worked and it never was a big problem. It's absolutely better than discard data and tear voice apart.

So the software should select appropriate delay for current network conditions and slowly change it if necessary.

You needed to know the dialling hack to force it via cable rather than satellite.

I think it was adding an extra digit to the dialled number - I have forgotten the details as it was a long time ago

That's what I meant by unusable. Web browsing[1] is unusably slow if you have these packet loss rates. Look at eg p6 graph here: https://netcraftsmen.com/wp-content/uploads/2014/12/20120410...

[1] and other non-interactive apps like social media

It may be the "jitter and other delays". They don't necessarily slow down browsing as much as packetloss. I think modern wifi causes a huge amount of it when other people in the same freq are using lots of bandwidth and it doesn't necessarily manifest as loss.

Right, if a packet gets delayed by 400ms, that's not going to break TCP, but you might as well consider it dropped for a real-time call.

If it's frequent (3-20pct), it's going to break your web browsing experience on TCP too because it had to be retransmitted. The original eventually appearing 400ms later doesn't make a difference.

Main threshold is if it is beyond RTT or not. But if it is 20pct frequent, there's 4% that are held up and then have their retransmission held up too.

Back to voice, many voice codecs will do forward error correction, but if jitter affects multiple packets in the sequence similarly instead of being independent that might not be enough.

I can't tell if this a April fool joke or real

Google isn’t doing April Fools this year.

Next step is to build speaker model on device and send it to peer during call setup

Not to sound negative, but what’s the point?

Does people really still use a Google-based communication program after all the ones Google has killed in the past? Is there an actual user base to reach using them?

Personally I’m avoiding them all and never investing in any of these again.

Also: We’re a few years down Duo still has fewer features than the other Google-owned IM-product it “replaced” ffs.

Rant over. (And yes, I’m still somewhat sore about that whole Gchat thing and how that went down.)

Google Pay, Youtube and Duo are very very popular in India. Try thinking global and not so US focussed for everything

> Does people really still use a Google-based communication program

Yes.

After years of trying to get my parents to use Skype, they have willingly adopted Duo.

I'm not sure if it's ease of use, or reliability, or one being associated with desktop and the other with mobile, but they call on Duo even if I say "Skype".

The thing about Google is that they're pretty R&D heavy. Making Duo into a killer video app isn't the point, it's to handle all the technical edge cases in a problem space and then apply that to every system at Google.

I use Google Fi for phone calls and it seems like it's fairly lossy, but they can take their improvements, such as this from Duo, and apply it using software to make it less noticeable to me, a user on a completely different product line.

Why would a potential customer want to invest time into learning a product that’s just a temporary research project?

Well, it's also to have a legitimate Facetime competitor, that's cross-platform and ready to roll for Android, and first-party.

I use them, because when they disappear I can quickly move to a new one. If I'm building a data "asset" then I am careful about using something that will be around for the long term but if it's just for an ephemeral video call I'll use Joe's until he goes out of business then I'll use Bob's.

Reply on news.ycombinator.com