WebRTC Virtual Interim

30 March 2020


Bernard, DomHM, Florent, Harald, Henrik, Jan-Ivar, Jianjun, SamDallstream, TimPanton, Youenn
Bernard, Harald, Jan-Ivar

Meeting minutes


WebRTC Features at risk

Bernard: a few unimplemented features not yet marked at risk
… 3 issues filed related to that
… first one is Issue 2496 - the voiceActivityFlag exposed in SSRC, not implemented anywhere
… any disagreement to marking it at risk?

Henrik: SGTM

Bernard: we have one unimplemented MTI per issue 2497, partialFramesLost
… should we remove it from the MTI list?

Jan-Ivar: no objection to unmark that one; will we get implementations for the other ones?

Henrik: they need to be moved from one dictionary to the other - they've been implemented, they just need to be moved into a different object

JIB: it's not clear to us yet how easy it will be to implement in Firefox; pointers to upstream webrtc.org hooks would help

Resolution: remove MTI marker on partialFramesLost

Bernard: last one is multiple DTLS certificates, not implemented anywhere

HTA: the goal was to help support signed certificates, which is completely unspecified

Dom: so if we remove support for it, the idea would be to say the spec only uses the first certificate in the list

TimP: wasn't the background of this support for multiple kind of certificates?

Bernard: with full support for DTLS 1.2, that's no longer relevant

Bernard: I'm hearing consensus on all of these

ISSUE-2495 When is negotiation complete?

JIB: this emerged while writing tests for WPT, but is applicable beyond testing
… "Perfect negotiation" is the pattern we recommend in the spec that helps abstract away the negotiation from the rest of the application logic
… having a negotiationended event would help avoid glare, simplify the logic
… the obvious approach to detect the end of negotiation is racy
… there are workaround, action-specific spin-tests (while loops)
… but that's bad, leading to timeouts
… I've also tried another workaround by dispatching my own negotiated event at the exact right time
… this is slightly better, but we can still miss cases
… can we do better? I have 3 proposals
… fire a negotiationcomplete from SRD(answer) if renegotiation isn't needed
… one downside is that subsequent actions may delay the event if further negotiations is needed in some edge cases
… Proposal B is a boolean attribute for negotiationneeded - needs careful implementation in relation to the negotiationneeded event
… it's also delayed by subsequent actions
… Proposal C: an attribute exposing a promise for negotiationcomplete
… it's better because it's not delayed by subsequent actions (by replacing promises as new negotiations get started)

Henrik: compared to proposal A?

JIB: imagine you call addTransceiver-1 & addTransceiver-2, you have to wait until addTransceiver-2 before the event fires (which you don't in proposal C)

Henrik: you can build your own logic if you care about partial negotiations - what you want to know in general is "am I done or not"?

HTA: I question the question - why should I care if negotiation is complete?
… what you have here is indeed a problem, but what the app cares about is whether the transceiver is connected to a live channel or not
… you don't have this problem with datachannels since you have an onopen event
… if we want to solve this at all (I would prefer not adding any API at this point), I think we should look at a signal on the transceiver availability

Bernard: don't you get that from our existing states, e.g. via the transports?

Harald: we have it with currentDirection, but without an event, it has to be polled

JIB: I think apps do need to know whether the transceiver is ready or not, and having that done with a timeout is not great

HTA: what I'm saying is what matters is the readiness of the transceiver, not the state of the negotiation
… if we want to add anything here, it should be a directionchange event to the transceiver

TimP: it could be done with proposal C which indicates "what" is complete (i.e. which transceiver is ready)
… otherwise, I agree you want to know what it is you got

JIB: you would get that via JS closure

Henrik: I think this is a "nice-to-have" - useful for testing & debugging; but I think it's a problem that can be solved with the existing API

JIB: I don't think this can be polyfilled, given that negotiationneeded is now queued
… negotiationneeded can be queued behind other operations in the PC

Henrik: you can detect this for each of your negotiated states by observing which changes are actually reflected (with different logic for each type of negotiation)
… this would be nicer, but I don't think it's needed

JIB: you mentioned setStreams - it cannot be observed locally
… another advantage of the promise is that it lets you determine if you're still on the same "negotiation train" by comparing promises

Youenn: it would be interesting to see if libraries built on top of PC are implementing that pattern
… this might be a good way to determine its appeal

Henrik: it would be great for debugging for sure, esp in the age of perfect negotiation

Youenn: so let's wait to see what apps adopting perfect negotiation before committing to this

Conclusion: keep for post 1.0 (in webrtc-extension?)

ISSUE 2502 When are effects of in-parallel stuff surfaced?

Henrik: the singaling/Transceiver states defined in JSEP and the API can't be the same to the cost of racy behavior
… which means the requirements imposed by JSEP on these states create ill-defined / inconsistent behaviors
… Proposals to address this: Proposal A: we make addTrack dependent only on WebRTC states, not JSEP states
… this is probably what the spec says, not what implementations do
… Proposal B: we make addTrack depend on a "JSEP transceiver", but would be racy and create implementation specific behaviors

JIB: I agree there is a race in JSEP
… JSEP was written without thinking about threads at all
… the problem is not really about whether we're in a JS thread or not
… we have to make copies of things

Henrik: my mental model is that WebRTC JS shallow objects refer to JSEP objects
… the only problem is with addTrack because of recycling of transceivers

JIB: the hygienic thing would be to copy state off from JSEP when looking at transceivers. Is that proposal A?

Henrik: it's implicit in proposal A

JIB: the only problem with that with your example on slide 17 - this would leave a hole e.g. in the context of perfect negotiation

Henrik: I think that's a better alternative than starting to meddle with internal JSEP objects
… the hole here is that if you're unlucky, you need another round of negotiation
… and in that situation, you would be in a racy scenario in the first place

HTA: the code of slide 17 is not compatible with perfect negotation

Henrik: I think proposal A is the only sane approach

HTA: this sounds ready for a pull request

JIB: I think the spec is currently racy given "JSEP in parallel" so it's more than an informative change

Resolution: getTransceivers() SHALL NOT be racy

Media Capture and Streams

Issue 671 new audio acquisition

Sam: Sam Dallstream, engineer at Microsoft
… this is a feature request / issue on the spec
… at the spec stands today, it is hard to differentiate streams meant for speech recognition vs communication
… the current implementations are geared towards communication, which sometimes is at odd with the needs for speech recognition
… e.g. in comms, adding noise can be good, but it hurts with speech recognition
… slide 22 and 23 shows the differences of needs between the two usages, extracted from ETSI specs
… the first proposal to address this would be a new constraint (e.g. "category") that allows to specify "default", "raw", "communication" "speechRecognition"
… it translates well to existing platforms: windows, iOS, Android have similar categories
… the problem is that it competes with content-hint in a confusing way - content-hint is centered around default constraints AND provide hints to consumer of streams
… whereas this one is setting optimization on the stream itself (e.g. levels of echo canceling)
… A second proposal is to modify the constraints to make them a bit more specific, and add a new hint to content-hint
… the advantage is that it fits the current content-hint draft, with more developer freedom
… but it may be hard to implement though
… Would like to hear if there is consensus on the need, and get a sense of the direction to fulfill it

Henrik: for clarification, for echoCancellation, it's not turning it off, it's tweaking it for speech recognition

Sam: right - right now echoCancellation it's a boolean (on or off)

HTA: but then how does it fit well well with the existing model?

Sam: I meant it's easier for API consumers, but you're right it conflicts with other constraints

Bernard: this is not about device selection here

JIB: indeed, most of this is done in software land in any case

Henrik: right, here it's more about disable/enabling feature

JIB: what's the use case that can't be done by gUM-ing & turn off echoCancellation, gainAutoControl, ambientNoise?

Bernard: it's not on & off

TimP: e.g. in speech interactions, you don't want the voice AI to hear itself

Sam: Alexa right now turns off everything and then adds their own optimization for speech recognition
… so this can already be done, but the idea is to allow built-in optimizations so that not everyone has to do their own thing

Youenn: do systems provide multiple echo canceller?
… I don't think you can do that in iOS

Sam: that's why the second proposal isn't as straightforward

Henrik: the advantage of these categories is that they vague enough that implementations can adjust depending on what the underlying platforms provide
… but then it's not clear exactly what the hint does

HTA: I would expect a lot of variability across platforms in any case

Henrik: as is the case for echoCancellation: true

HTA: indeed (as the multiple changes of the impl in Chrome show)

Henrik: it sounds like it is hard-enough to describe, implementation-specific enough that it should be a hint

JIB: I think that's fair to say that the audio constraints have been targeted a the communications use case
… not sure how much commitment there is for the purpose of speech recognition

Sam: right

Henrik: with interop in mind, echoCancellation: true worked because everyone did their best job at solving it, not doing it the same thing
… to get that done with this new category, we would need the same level of commitment and interest from browser vendors
… the alternative is turning everything off and doing post process in WebAudio/WASM

TimP: another category beyond comm, speech-rec here is broadcast
… it shouldn't be a two-states switch

JIB: anything here that couldn't be solved with WebAudio / AudioWorklets

Sam: I would need to take another look at that one

HTA: you would still need a "raw" mode

Youenn: maybe also look at existing open source implementation of ambient noise and whether they share some rough parameters

Sam: it sounds like leaning towards 2nd proposal

Dom: maybe first also determine what can be done in user land already with Web Audio / Web Assembly
… if this is already doable there, then maybe we should gain experience with libraries first

HTA: given we already have a collection of hints in content-hint that have been found useful, it's kind of easy to add it there

Bernard: would this applies up to gUM?

HTA: yes, that's already how it works

JIB: if we're thinking adding a new hint, we may need new constraints specific to speech-recognition

[discussion around feature detection for content-hints]

ISSUE 639 Enforcing user gesture for gUM

Youenn: powerful APIs are nowadays bound to user gesture
… if we were designing gUM today, it would be as well
… but that's not Web compatible to change now
… can we create the conditions to push Web apps to migrate to that model
… PR 666 proposes to require user gesture to grant access without a prompt
… I've looked at a few Web sites; whereby.com works with the restrictions on
… it wouldn't work in Hangout or Meet
… Interested in feedback on the approach and availability to help with webrtc app developers outreach

Youenn: the end goal would be that calling gUM without user gesture should be rejected
… user gesture is currently an implementation-dependent heuristic - this is being worked on

Henrik: I think we would need it to be better defined
… it is also linked to 'user-chooses'

Youenn: the situation is very similar to getDisplayMedia where Safari applies the user gesture restriction
… it could be the same with gUM

JIB: I like the direction of this; we could describe it as privacy & security issue
… with feature-policy, there is a privacy escalation pb through navigation
… jsfiddle allowed all feature policies, so from my site I could have navigated to my jsfiddle, got priviledged there before navigating back with an iframe
… so that sounds like an important fix
… the prompting fallback sounds interesting
… denying on page load might be harder to reach
… it's not clear that same-origin navigation should be blocked

Youenn: user gesture definition is still a heuristic, these could fit into that implementation freedom

HTA: how much legitimate usage would we break?
… before progressing this, we should have a deployed browser with a counter to detect with/without user gesture

Youenn: Webex and Hangout call it on pageload, so that would make the counter very high

HTA: so will someone get data?

Youenn: I don't think Safari can do this; would be happy if someone can do this
… I can reach to top Web site developers

HTA: would anyone at Mozilla interested in collecting this data?

JIB: based on our user gesture algorithm? I'll look, but can't quite commit resources to this at the moment

Conclusion: more info needed

Next meeting

HTA: probably in April / May

<dom> s/Topic: Issue-2495/SubTopic: Issue-2495

<dom> s/Topic: Issue 2502/SubTopic: Issue 2502

Summary of resolutions

  1. remove MTI marker on partialFramesLost
  2. getTransceivers() SHALL NOT be racy
Minutes manually created (not a transcript), formatted by scribe.perl version 114 (Tue Mar 17 13:45:45 2020 UTC).