Meeting minutes
Agenda
Chris: We will continue the discussion about AI and Media Entertainment today.
… next TPAC meeting is in Dublin, will be looking for agenda topics (later)
AI media use cases
Chris: Mr. Zheng and Song presented AVS standardization use cases, like 2D 3D image converison, volumetric use cases etc.
Song: AVS in China use cases were shared, and a demo, but we only had limited time to discuss so I'd like to continue that conversation.
… We could break down specific features / characteristics of the use cases and categorize in different levels. Then describe impact on MEIG.
Chris: Sounds good, and we should gather input from more members here on what they are interested in (AI usecases as they apply to media)
… so we have as broad a set of usecases as possible. We are interested to capture your usecases.
Louay: MOQ could be a topic for the TPAC meeting as well.
… is related to AI for low latency streaming, if you have AI chatbots or voice assistants, or avatars, that are generated in the backend and rendered in the browser
… that would make MOQ relevant as well in the AI context.
Chris: Let's capture the AVS categorization.
https://
Song: When categorizing use cases, I am thinking end to end use cases. Codecs are relevant, and transportation.
… I believe that what Louay mentioned is also relevant for discussion, would be appreciated if transportation could be included.
… GitHub has the original information from Dom on the Roadmap for AI Technologies on the Web, the intention is to see if there are changes in priority
https://
Song: With Chris' response about labelling, I started. My understanding of the domain is that there are several groups of AI efforts.
… C2PA is in the first area; relevant for content authored by humans and assisted by AI; or by AI alone
… these impacts can be group 1: text based content generation.
… Beyond that, for MEIG, more complicated impacts on WebCodecs, A/V, WebXR etc., this can be second group.
… Three use cases: next-gen codecs driven by neural networks (UC1); enhanced experience with immersive experience (volumetric, super resolution, ...) (UC2)
Dimitri: do you know who is driving the work on MPEG DCVC?
Song: Tech people from Microsoft Research Center AFAIK; within MPEG, I observed that some of attendees talk about DCVC but have not started official standardisation process.
… The third sub-usecase is multimodal agents combined with text and A/V on the web (UC3).
… If we add the content labelling scenario, that would be four use cases.
Chris: We may have more to add from input in this group
Song: In general, I believe that UC1 and UC2 have similar foundations, in that they are upgrades of existing codecs to/with AI, or enhancements of the user experience
… They have features in common, for example raw data for rendering is same/similar.
… WebGPU and webNN enhance the codecs
… Volumetric could lead another direction: higher data rates, higher demands on bandwidth, lower latency streaming for WebTransport
… For UC3, they are the most emerging use cases; people use ChatGPT, Claude, Doubao, Qwen through web platform. In the future, these agents will be upgrade from text based input to multimodal.
… The simultaneous inference from vision and audio could demand higher requirements
… realtime requires WebTransport for low latency
… Is there a chance for optimization or a need to integrate AI labelling?
… Key output transforms web media from codec to rendering to agentic interaction
… WebGPU and webNN/webTransport are foundational
Chris: Louay, what are your thoughts about MOQ and WebTransport?
Louay: MOQ/WebTransport is the basic layer for low latency. People today use WebRTC
… I can provide quick intro into MOQ and relation to WebRTC and the browser, summarize use cases (media distribution, data, AI voice assistance, production)
… at least WebTransport is required by MOQ which is built on QUIC. WebTransport is also important for realtime but can also be used with MSE but loses some control
Nigel: Another use case. I had an action to put together a TF proposal on AI usage in subtitle and captions. I still have that action to do.
Chris: We should add that to the use cases on this slide.
… Nigel's use case is about using AI for captions and generating descriptive metadata like emotion or audio loudness levels that can enrich the description of what is happening in A/V
Song: Responding to previous topic about WebTransport, we had this before in other working groups. When we talk about WebTransport we talk about a better user experience
… it could be important foundation for multimodal agents. Even though the topic is similar to earlier, it is a research item in AI
Chris: Agree, this is foundational.
… At this stage, is it enough that we have identified it? Louay, a future look at it would be welcome, let's follow up.
<Zakim> nigel, you wanted to mention my action on new requirements for subtitles enabled by AI (can be later)
<kaz> Smart Voice Agents Workshop Report - 25-25 Feb. 2026
Kaz: Thank you for the information provided, these are important use cases. In February there was the W3C Workshop on Smart Voice Agents.
https://
https://
Kaz: One of the topics was relevant for voice technologists. I think it would make sense to have another workshop on Media & AI including various stakeholders across the world.
… The target area you're describing is getting broader, so makes sense to have more stakeholders including browser vendors, and AI vendors. Thoughts?
Song: I got involved in the voice agent workshop; you are correct that for voice agents we organized this workshop. Could be useful to come up with a similar analysis or report, setting different targets or milestones. The use case characterization will be useful.
Kaz: The main topic for the Voice Agents workshop was voice agents, while our own interest is "Media and AI" including AI bots and AI agents in general. So our target is even broader. A smart agent workshop would make sense.
Song: Will review the output for the voice agent workshop and check whether results can be used for further use cases.
kaz: My main point is thinking about our own workshop about media and AI.
… And we already have three use cases for the workshop.
Chris: This is something that we could plan to do, let's consider it.
Rob: What ties into use case 4 (content labeling), looing at this from a privacy/security angle, if you think about a still image on the web,
… we see things in it but which are difficult for machines to extract. But this is becoming easier. Now, if I can recognize something then so can the AI.
… If all this data becomes available, then agents can go and label images with who is in them; or the audio. This becomes a dystopian scenario where one can search for me and the engine comes up with all locations I've ever been.
… This is something we should guard against. There are already robots.txt files that guard some if this; we should think about something similar for AI.
… My premise is that if I as a human can do it so can AI (but at a larger scale).
… So we need to think about what data can be aggregated from all of this?
… This relates to the content labelling mentioned earlier.
Chris: Ah, the labelling as intended here was meant as signal that content was AI generated; regulations are coming up in the EU and US that mandate this.
… There is a spectrum from AI assisted but with human oversight, to completely AI- generated. The C2PA technology generates signatures to make that signalling metadata tamper evident -- any editing would be visible and can be made traceable.
… You end up with a content provenance log. The big driver for the development of that is generative AI.
… The labeling that you are talking about is a bit different, using machine learning to understand media content and generate metadata.
Rob: Yes, this is something that humans can do and so can AI. There could be an issue with, say, 1000 pictures of a person, and one can mine for commonalities.
… How are we protecting against this?
Chris: Within my organisation, we have a large archive of media. A usecase for us would be to find all clips of a politician, for example. And it is exactly as you describe, use AI to discover what's in the media. Same with voice, find who speaks at any one time.
… What I like about your description is it is across the entire web.
… At the moment, AI is trained mostly on text, but how do the concerns of privacy and protecting people's interests change when it's also A/V?
Rob: If you regard the internet as a library, finding something by text is possible using search engines. Once this is available for images/audio, more use cases become available through aggregation: for example, you can discover who met whom, and where and when?
… AI can discover these hidden connections. And every member of the public can make these inferences.
Chris: There is also a usecase that you mentioned, around what content should be accessible to the AI?
… This is worth capturing.
… And there is active work on that, e.g., in AI Preferences WG at IETF.
Rob: The use case could be called "Rapid Access to media"; WebVMT ties into this
… An example of this: OGSC testbed was just finished.
… The dashcam in my car has an accelerometer; could look for a spike which would indicate a collision. Horizontal spike: collision; vertical spike: pothole. Tie this to location and you find potholes and speed bumps
… people can look at this demo.
<RobSmith> OGC Testbed-21 Dash Cam Analysis Video: https://
Chris: This is an example of annotating interesting points in time in the video, worth capturing as another use case.
… Does anyone else have interesting use cases?
Louay: Capturing emotion is another example of multimodal analysis. LLM systems and other components of TTS producing audio lose emotion, currently. Maybe LLMs can be asked to provide emotion as part of the output?
… Putting emotion into the voice agent is important for many use cases. You need emotion in audio, and in the avatar facial expression. This is another use case.
Chris: We have some work to do to capture the use cases.
Song: I'll go through the notes and complete the use cases.
Chris: We can then build on this.
<RobSmith> I summarised my use case in a comment in issue#123: w3c/
TPAC 2026 planning
Chris: One more thing. I updated the GitHub issue with our meeting times at TPAC. Please think about what topics we want to cover in the times that we have.
… In September we need to pull these together and form a plan.
w3c/
Next meeting
Chris: Our next meeting is on August 4th. In the meantime, please find the GitHub issue and add your use cases: w3c/