Google Announces Lyra V2 Low Bit-Rate Voice Codec

[anarki2]

That's cool and all, but make these part of BT already. Without that, all these codecs are utterly worthless.

Also, Lyra 2 is a stupid name for this, I immediately think of crypto.

[Alliancemd]

The SoundStream version showcased last year was able to handle music as well, wonder if Lyra V2 has better success with that. Someone on Discord needs to re-encode 8MB Shrek....

The processing latency was reduced from 100ms, to 20ms, so it should do even better now

[Jedibeeftrix]

I agree on the point "interesting technology", but i doubt it will become the data equivalent of CB radio.

The whole thing is essentially owned by one company called "Semtech", they made it a patent minefield and straight out patented the way the data is modulated / transmitted and charge a lot of royalities for it. That´s why the tranceiver chips are 7$+ typically, as most of this is royalities.

All the different parts of LoRa are straight forward and already well known since years, but they patented the whole solution / charge royalities for it..

Guess we need an Open Standard with the same Range / Data Rate / Mesh things without the royalities to have something that really takes off.

Ooh, didn't realise this.

But if you look at the life-cycle of the industry - where device lifetime alone is in the 10yrs unattended range - then much like CB we'll still be using this in 50 years.
When were these patents filed?
In the UK the average patent last only 15 yrs i think (unlike copyright where you're looking at north of 60 yrs).

[milkylainen]

3.2kbit/sec that would enable voicechat over things like LoraWAN / could have applications for long range voice transmission via RF with very low transmit powers, as the symbol rate could be very low / the receiver could differenitate the data from background noise via long integration times.

I think people need to realize that AI/ML symbol replacement is not really compression in the classical sense.
Unless you prefer to call speech-to-text a lossy form of audio compression too?
This type of data handling moves the result into a new domain.

First off you need a very large symbol library. Secondly, I highly doubt that this type of analysis can be made on low power devices (in software, yet).
The result might be viable for an application and it definitely has its uses... but it's not compression.

I mean, I could transfer the whole audio as classically compressed very low bitrate phonetics and have a computer voice read it back.
I would definitely win as far as bitrate goes. It's sort of the same thing.

[CommunityMember]

it would be interesting to see comparisons between Lyra V2 and Codec 2

As I recall, one of the challenges is while Codec 2 is itself patent free, the Wavenet decoder (an important part of many of the implementations) is patented. Presumably Lyra is fully patent free. And (just as with AV1) a solution without IP encumbrances is going to be preferred in the open source world.

[satadru]

3.2kbit/sec that would enable voicechat over things like LoraWAN / could have applications for long range voice transmission via RF with very low transmit powers, as the symbol rate could be very low / the receiver could differenitate the data from background noise via long integration times.

Exactly this. I'd love to find some low power LoraWAN hardware which could also be powerful enough to handle Lyra and some simple voice recognition. Let's get some open source Star Trek communicators working!

[RealNC]

If Lyra's aim is to compete with codecs that want to achieve audio transparency (not being able to distinguish between original and compressed audio,) then the comparison to Opus is fair, I think.

[Spacefish]

Lorawan is an interesting technology, can almost see it becoming the data equivalent of CB radio.

I agree on the point "interesting technology", but i doubt it will become the data equivalent of CB radio.

The whole thing is essentially owned by one company called "Semtech", they made it a patent minefield and straight out patented the way the data is modulated / transmitted and charge a lot of royalities for it. That´s why the tranceiver chips are 7$+ typically, as most of this is royalities.

All the different parts of LoRa are straight forward and already well known since years, but they patented the whole solution / charge royalities for it..

Guess we need an Open Standard with the same Range / Data Rate / Mesh things without the royalities to have something that really takes off.

[Jedibeeftrix]

3.2kbit/sec that would enable voicechat over things like LoraWAN / could have applications for long range voice transmission via RF with very low transmit powers, as the symbol rate could be very low / the receiver could differenitate the data from background noise via long integration times.

My first thought too.

Lorawan is an interesting technology, can almost see it becoming the data equivalent of CB radio.

[rshpount]

I assume Michael meant video chat at 56kbps, which is possible with AV1 and Lyra combination.

The main competition for Lyra is MS Satin. All Google is trying to do is to keep Google Hangouts competitive with Teams.

ML is not running during Lyra encoding or decoding process. Lyra is using a pre-trained neural net. It is roughly as CPU efficient as traditional LPC codecs.