Server-side audio filters and latency control

[erkinalp]

Currently, there is very little latency control in Discord's voice channels. While this is usable in most settings, this does not work in the arts study groups and serious conferencing use cases. I mean latency control as in delay-based compensation of latencies such that everyone in the channel would hear exactly the same stream with no resampling taking place.

Also to mention, there are currently no server-side audio filtering that would allow moderating a voice channel for all of its users. If such ability existed, we could easily remove curse words, foreign language speech, or copyrighted content out of those voice chats.

I opened a single ticket for both features as they require similar architectural and API changes.

[hwangcha-x-bts]

u ok?

[TwilightZebby]

...I don't think you realise just how much more difficult it is for automated audio moderation vs text moderation.

[yoshi32]

How about it does my dishes too?

[Mineinjava]

Would you mind explaining how this would work?

[erkinalp]

Would the filtering be done client or server side? This would be expensive stuff!

Server side, like ToxMod does.

[splatterxl]

Wow, looks good! 😁 I'm looking forward to this being available to servers as a much-needed moderation tool

[NovaFox161]

Toxmod is not a filter. It's simply a tool to allow humans to review after the fact.

Please, prototype a realtime filter. Honestly, if you pull it off, just sell it to the highest bidder because that's some high tech stuff

[Mineinjava]

I know that voice moderation is simple--it can already be done by having bots "listen" to other users and run it through a speech to text engine. I was wondering about how you envision the latency control to work.

[erkinalp]

Latency control is explained in the main post: by adaptive delay-based input.

[RGBCube]

its actually really easy to do, just record someone saying bad words and check the audio stream if it matches the bytes in it, no ai needed at all!

[TwilightZebby]

If it really was that simple, YouTube's and Twitch's Copyright/DMCA systems wouldn't be having so many false-positives 😂
(I raise the classic example of a band getting banned from Twitch for playing their own music on their own live stream)

[RGBCube]

If it really was that simple, YouTube's and Twitch's Copyright/DMCA systems wouldn't be having so many false-positives 😂
(I raise the classic example of a band getting banned from Twitch for playing their own music on their own live stream)

i know lol this is sarcasm

[Katzen48]

You know...for the moderation stuff...
The examples you stated contain three different kind of filters: bad words, foreign languages and copyrighted content.
Each of them requires a different model for an AI to process it.

Bad words = NLP model for the specific language, bad words database and intent detection
Foreign languages = model to detect the language, someone is speaking
Copyrighted content = model trained with all copyrighted righted content

Something like this is so expensive and time-consuming, that even companies like Google, Facebook, Amazon and Microsoft are having their troubles, creating even one of those models in a reasonable time.
Where should Discord get the budget or the man power for that from?

Another thing is, that stuff like "bad words" are very regional.
Stuff that is banned in my country, does not have to be banned in yours.

Last but not least, this also sounds like censorship and undemocratic.
Someone decides for me, what I may say or play on a Discord before that information even got delivered to the other people.
Data science is far away from delivering a perfect filter, that only filters what it should filter. There is a statistical inaccuracy in this.

I believe, there is no real use case for something like this on a platform, that stands for open discussion.

[erkinalp]

Foreign language detection + bad words is really the same filter as bad words vary depending on the language. Something that stands for a curse word in one language could be an essential verb in another.