Autor: Redaktion

Artificial Intelligence is For Real (3 / 3)

No, AI isn´t just a hype.

In considering archive, our session was looking at existing content collections, rather than new ones that would already have benefited from the application of AI processes earlier in the production chain. 
It is often assumed there is inherent value tied up in all archives – and that the only barrier to the release of that value is the difficulty and cost of making the content accessible. This is not necessarily the case. Not all archived material is of interest to consumers or has any reuse value. However no one knows the value of any archive until all the content within it can at least be identified; and very few archive catalogues are sufficiently authoritative to provide this information.
The greatest opportunity represented by AI therefore is around sound and picture recognition and transcription, so that an index can be created once the content has been digitised. This still poses cost challenges however, and the return on investment is inherently difficult to predict if the amount of valuable content in an archive is unknown. So the focus for our experts was as much on the ability of AI to spot-buy low cost use of cloud resources to make the analysis more cost effective, as on content analysis itself. 
Once such indexing has been achieved then AI does offer the potential to select only the archive material that is likely to have value, and even to automatically compile that content into collections, complete with tags. actual 

Some auto generation of content from archive collections is beginning to occur, as is the analysis of audio as a relatively low cost means of assessing whether the content to which it is attached is of value
Archive SCORES
[table][tr][td]Potential [/td][td]AI Type [/td][/tr][tr][td]Content selection  [/td][td]B2 [/td][/tr][tr][td]Compilation  [/td][td]B2 [/td][/tr][tr][td]Automated metadata tagging [/td][td]B2 [/td][/tr][tr][td]Auto hierarchical storage management       [/td][td]A1 [/td][/tr][tr][td]Spot instances [/td][td]A1 [/td][/tr][/table] 
[table][tr][td]Efficiency [/td][td]4 [/td][/tr][tr][td]Opportunity      [/td][td]4 [/td][/tr][tr][td]Experience [/td][td]3 [/td][/tr][tr][td]Total [/td][td]11   [/td][/tr][/table] 
[table][tr][td]Actual [/td][td]AI Type [/td][/tr][tr][td]Auto content generation  [/td][td]B1 [/td][/tr][tr][td]Auto logging of audio to assess archive monetisation value       [/td][td]B1 [/td][/tr][tr][td]Fan tagging of video archive  [/td][td]B2 [/td][/tr][tr][td]MATURITY SCORE [/td][td]3 [/td][/tr][/table] 
Phase 2: Fulfil
As more and more content has global distribution via the Internet, and as there is increasing awareness of the need to deliver content that responds to the accessibility requirements of consumers, versioning has become a crucial – and hugely complex – element of content fulfilment. 
The attempt to meet these demands through manual processes is labour intensive, expensive and difficult. In many respects AI is ideally suited to tasks such as this, which are repetitive, detailed and occur at scale. 
But on the other hand high quality versioning is all about cultural nuance – and this is where computers struggle. There is, nonetheless, a genuine business need for any technology that reduces the cost and complexity of versioning and that supplies more appropriate content to more people in more territories. This means AI seems certain to play an increasingly important role in versioning.
There are opportunities for AI to provide auto transcription for subtitles, for search purposes and to provide data required for targeted advertising. It is conceivable that AI could transform the dubbing process, using combinations of speech to text, language detection, translation and speech synthesis. It could similarly improve audio description; and it is not unimaginable that one day there could be avatar-based on-screen signing. 
More prosaically AI could be used to detect, and sometimes replace, branding and adverts – whether for compliance or viewer preference reasons, or because of commercial agreements and opportunities.
The reality of how AI is impacting versioning today is considerably more modest – but already beginning to deliver some benefits. In areas such as transcription, there is still the need for human review. But even a transcription process with 80% accuracy delivers benefit and reduces the human workload. 
The ability to detect and replace signage already exists, and is in use around sports content, where stadia provide fixed and predictable environments.
Versioning SCORES
[table][tr][td]Potential [/td][td]AI Type [/td][/tr][tr][td]Automated dubbing  [/td][td]B2 [/td][/tr][tr][td]Automated subtitling  [/td][td]B2 [/td][/tr][tr][td]Automated audio transcription       [/td][td]B2 [/td][/tr][tr][td]Automated audio description  [/td][td]B2 [/td][/tr][tr][td]Avatar based signing   [/td][td]B1 [/td][/tr][tr][td]Brand detection [/td][td]B2 [/td][/tr][/table] 
[table][tr][td]Efficiency [/td][td]4 [/td][/tr][tr][td]Opportunity      [/td][td]3 [/td][/tr][tr][td]Experience [/td][td]4 [/td][/tr][tr][td]Total [/td][td]11 [/td][/tr][/table] 
[table][tr][td]Actual [/td][td]AI Type [/td][/tr][tr][td]Auto language detection  [/td][td]B1 [/td][/tr][tr][td]Channel logo rebranding  [/td][td]B1 [/td][/tr][tr][td]Credits detection  [/td][td]B1 [/td][/tr][tr][td]Transcription to enable search for journalists  [/td][td]B2 [/td][/tr][tr][td]Content protection – constraining material visible to dubbing artists      [/td][td]A1 [/td][/tr][tr][td]MATURITY SCORE [/td][td]4 [/td][/tr][/table] 
Compliance and Delivery
Compliance and delivery is very closely linked to versioning, since these processes will need to be applied to each version. The processes around ensuring that content is compliant, from regulatory, quality, technical and contractual points of view, are some of the most fundamental, costly and meticulous in the world of content creation. Overall, quality is on the increase, and many content providers stake their reputation on the delivery of highly compliant content. 
The benefit of the application of AI to these processes will be felt more strongly – or at least more consciously – by content providers than consumers. Consumers are largely unaware of the effort that goes into making content compliant, until it goes wrong – at which moment they are likely to be highly aware, and their view of the content provider is likely to be significantly damaged. 
This means that the stakes for any automated processes in this area are high. Reliability is key. It would take considerable technical maturity before most providers would be comfortable about relying solely on AI, without human review.
The application of AI to image and audio recognition has enormous potential to identify compliance issues such as smoking, nudity, culturally sensitive material, swearing, and so on. It could also, potentially, and in certain circumstances, remove such content to create a redacted version.
Basic quality control (QC) processes still represent a significant overhead for many content providers, especially those providing premium quality content. It is normal for QC to be repeated a number of times at various stages in the fulfilment process. The cost saving – and time saving if high quality AI-led QC could be carried out in less than real time – would be significant. However QC currently represents a small proportion of the total cost of premium content, while the reputational damage of QC failures can be significant. So the benefits of full auto QC will only be realised if reliability is extremely high.
Auto QC already takes place to a degree, and is delivering real benefits to some content providers. Currently however, manual review of automated QC outputs can be time consuming, so evolutionary steps to reduce “false positives” may be what we see next. Automated QC is still supplemented by human review at some stage in the delivery process – often referred to as editorial or ‘eyeball’ QC. The benefits are around cost rather than quality or consumer experience, since the quality of current human led QC is very high.
There are already use cases around auto recognition of faces, notably for the identification of celebrities and politicians – both from pre-recorded and live content.
Online platforms are already using auto detection of content for reasons of content moderation and copyright monitoring. There is also some automatic generation of subtitles.
Compliance and Delivery SCORES
[table][tr][td]Potential [/td][td]AI Type [/td][/tr][tr][td]Quality Control  [/td][td]B2 [/td][/tr][tr][td]Compliance and redaction       [/td][td]B1 [/td][/tr][tr][td]Bad language detection  [/td][td]B2 [/td][/tr][tr][td]Logo detection [/td][td]B2 [/td][/tr][/table] 
[table][tr][td]Efficiency [/td][td]4 [/td][/tr][tr][td]Opportunity      [/td][td]3 [/td][/tr][tr][td]Experience [/td][td]2 [/td][/tr][tr][td]Total [/td][td]9 [/td][/tr][/table] 
[table][tr][td]Actual [/td][td]AI Type [/td][/tr][tr][td]Caption alignment and compliance  [/td][td]B2 [/td][/tr][tr][td]Automated online subtitling  [/td][td]B2 [/td][/tr][tr][td]Facial recognition of celebrities  [/td][td]B2 [/td][/tr][tr][td]Auto Quality Control  [/td][td]B1 [/td][/tr][tr][td]Online content flagging for piracy, moderation and copyright      [/td][td]B2 [/td][/tr][tr][td]MATURITY SCORE [/td][td]3 [/td][/tr][/table] 
Historically, programming schedules in broadcasting have been decided well in advance of transmission, and the costs associated with maintaining the capability to make last minute changes to those schedules – because of national events or live programme overruns – have been enormous. Indeed, the huge costs associated with playout contracts have to some degree been related to ensuring contingencies for events that rarely happen.
But if linear programme scheduling appears rigid and crude, online scheduling of content has been similarly blunt. Here the issue is not about the scheduling of channels but rather the scheduling of adjacent content: what plays after or around a particular piece of content. Inappropriate adjacencies can generate significant controversy.
So although scheduling may appear to be a dull and mechanical process that would be easily susceptible to automation, in reality it is highly complex, requiring both responsiveness and sensitivity. And as advertisers become more and more aware of the benefits of scheduling video advertising in the optimum context – not only from the point of view of the relevance of adjacent content, but also with regard to mood and emotion – the ability to deliver sophisticated scheduling now carries greater commercial benefits than ever before.
The greatest opportunities in the application of AI to scheduling relate to advertising. There is the potential for deep learning systems to provide highly contextually relevant advertising in real time. Such targeting could also occur at the personalised level, and could respond not only to the tone and type of adjacent content but even to the habits, preferences and mood of the individual consumer. It is not an exaggeration to say that the future of video advertising lies with the development of such AI-led capabilities.
The commercial benefits are therefore considerable. But so are the creative and experiential benefits, since consumers will be served content that is more relevant to them, and they will be less frequently irritated by content they find irrelevant or, in the worst cases, offensive.
These consumer experience benefits extend beyond contextual advertising. Intelligent scheduling is the ultimate form of recommendation: if content providers appear to intuit correctly what a consumer would like to see next, then engagement will be greatly increased and the viewing experience will be enhanced.
Automation is slowly making its way into linear scheduling. Some of the basic, repetitive human processes around the scheduling of content – especially for non-live and niche channels – is already beginning to be automated. Spreadsheet based schedules are beginning to be ingested into automated scheduling systems.
In the online world, scheduling covers a few different tasks: planning availability windows for VOD content of course, but one could also include auto-play functionality that mimics a continual linear channel by playing VOD assets chosen for the user. While the former use-case is most often deployed using simple rules-based automation, the smartest video platforms are using deep learning for targeted auto‑play.
Scheduling SCORES
[table][tr][td]Potential [/td][td]AI Type [/td][/tr][tr][td]Automated metadata delivery  [/td][td]A1 [/td][/tr][tr][td]Automated break construction       [/td][td]B1 [/td][/tr][tr][td]Contextual advertising  [/td][td]B2 [/td][/tr][tr][td]Automated scheduling [/td][td]B2 [/td][/tr][/table] 
[table][tr][td]Efficiency [/td][td]3 [/td][/tr][tr][td]Opportunity      [/td][td]4 [/td][/tr][tr][td]Experience [/td][td]4 [/td][/tr][tr][td]Total [/td][td]11 [/td][/tr][/table] 
[table][tr][td]Actual [/td][td]AI Type [/td][/tr][tr][td]RPA-led entry of schedules to channel systems       [/td][td]A2 [/td][/tr][tr][td]Rules based auto-slotting  [/td][td]A1 [/td][/tr][tr][td]Linear schedule pre-checks  [/td][td]A1 [/td][/tr][tr][td]Automated advert insertion in VOD  [/td][td]A1 [/td][/tr][tr][td]VOD scheduling and autoplay [/td][td]B2 [/td][/tr][tr][td]MATURITY SCORE [/td][td]3 [/td][/tr][/table] 
Distribution (Linear and Online)
The biggest challenge for distribution in a world of both linear transmission and online video, is capacity and bandwidth.
AI-led encoding, including Statistical Multiplexing (StatMUX) – a way of optimising data streams to reduce online video bitrates or to maximise the number of channels broadcast in a fixed bandwidth – helps to deal with the need to meet this challenge.
One of the anxieties about the rapid shift to online viewing is that the content delivery networks will struggle to deal with the demand on those networks, particularly if consumers are trying to access high definition content for major, global events. There is therefore a genuine business need for the application of AI to manage capacity and optimise data traffic.
From a consumer point of view, such improvements will be experienced as an improved quality of service, which is likely to drive online and mobile viewing still further – possibly hastening the decline of linear programming, and certainly ensuring there is a never-ending need for technology innovation around the management of load in the distribution process.
The opportunity in this area is limited and specific. AI-led encoding and StatMUXing will deliver marginal improvements in picture quality, and greater reliability of service. Predictive fault finding will also enhance reliability. If fault finding is backed by AI led fault repair (‘self-healing’ systems), then there will also be efficiency savings.
Similarly the use of AI in capacity planning – predicting viewing patterns based on time, content, habits and so on – will enable content delivery networks (CDNs) to prepare more effectively for peaks in load.
This is one of the more mature areas for AI technologies. AI powered encoding, capacity planning and fault finding already exist as well as automated file delivery.
Distribution SCORES
[table][tr][td]Potential [/td][td]AI Type [/td][/tr][tr][td]AI powered encoding       [/td][td]B2 [/td][/tr][tr][td]Capacity planning [/td][td]B2 [/td][/tr][/table] 
[table][tr][td]Efficiency [/td][td]2 [/td][/tr][tr][td]Opportunity      [/td][td]2 [/td][/tr][tr][td]Experience [/td][td]2 [/td][/tr][tr][td]Total [/td][td]6 [/td][/tr][/table] 
[table][tr][td]Actual [/td][td]AI Type [/td][/tr][tr][td]Predictive fault finding  [/td][td]B2  [/td][/tr][tr][td]AI powered encoding  [/td][td]B1 [/td][/tr][tr][td]Pre-caching content to CDNs or devices based on predicted demand       [/td][td]B2 [/td][/tr][tr][td]Automated file delivery [/td][td]A1 [/td][/tr][tr][td]MATURITY SCORE [/td][td]4 [/td][/tr][/table]

Related Articles: