Warner Bros. Discovery (WBD) ended 2025 with 131.6m direct-to-consumer streaming subscribers worldwide. As HBO Max expands abroad, the company is becoming a global platform with an increasingly diverse audience.
This growth raises a critical question for the organisation: how can it scale globally while delivering localised content that resonates across markets? Answering it will be essential to unlocking the full potential of its global subscriber base.
Research suggests that 76% of millennials and Gen Z viewers have an appetite for international content1, and Hollywood’s library is well placed to meet that demand.

76% of millennials and Gen Z viewers have an appetite for international content.
Reaching that audience, however, requires one crucial capability: localisation. With just 15% of its catalogue available in languages other than English on HBO Max and its additional platforms, WBD saw an opportunity to expand its international reach dramatically—and turned to AI to deliver the speed and scale necessary, while preserving quality.
The company was already using machine learning extensively across its platforms, from content discovery to monetisation. But localisation was a new opportunity. Rather than treat it as an afterthought, WBD began constructing an AI-powered pipeline capable of processing petabytes of audio and video with unprecedented throughput and accuracy.
“Using AI to make localisation via captioning more affordable, smarter and faster is not just about efficiency,” says Deepna Devkar, senior vice president of machine learning and artificial intelligence for global streaming at WBD. “It also has the potential to expand access and cultural reach. It’s about democratising our rich catalogue and enabling as many fans as possible around the world to connect with our content in their language of choice.” The effort began in late 2024.

Designing an AI pipeline
WBD had already assembled a small but high-calibre engineering team, yet leadership recognised the need for additional strategic partnerships. Enter Google. “We’ve always seen our partnership with Google Cloud Consulting as a way to bootstrap and accelerate, not outsource,” Dr Devkar says. “The goal was to leverage advanced technology, parallelise our efforts and gain momentum on multiple fronts.”
The aim extended beyond improving captioning to building a scalable foundation for long-term growth. While WBD engineers refined their captioning models and core platform, Google helped prototype the translation pipeline. Rob Birdsong, a vice president at Google Cloud Consulting, the team responsible for partnering with WBD on the delivery, likens the dynamic to personal training: “They’ve been in the gym, they’ve been working out, and they just needed someone to help them get to the next level.”
Building the pipeline presented challenges. High-quality training data for “diarization,” the task of determining who is speaking when, was scarce. “We quickly realised that we couldn’t just rely on public datasets,” Dr Devkar recalls. “They didn’t reflect the nuances of our content.” WBD established dedicated labelling operations to generate its own training data.
Integration was another pain point. Content traverses multiple supply chains, each with distinct asset-management systems. Early attempts to orchestrate everything through existing infrastructure proved too rigid, prompting a shift towards robust APIs—what Dr Devkar calls “a key architectural pivot” that delivered long-term scalability and interoperability.
Even with better data and infrastructure, off-the-shelf tools fell short. Speech-to-text and transcription software struggled with media-specific quirks such as background music, songs and overlapping dialogue. “Models like Whisper and Gemini are trained on very different types of audio and video data,” Dr Devkar explains. “But media content brings unique challenges.”

The team had to fine-tune its own models on proprietary datasets reflecting broadcast conditions, including music, multiple speakers and specialised vocabulary. A key part of this effort involved training customised, lightweight model adapters tailored to individual networks. Rather than rely on a single general model, this approach helps the system recognise network-specific language, whether that is culinary terms on a food channel or technical jargon on an automotive network.
Alongside retraining, WBD’s team also improved how audio is prepared and routed before it reaches the transcription model. Audio is now intelligently segmented and directed to the most suitable transcription method based on its length and complexity, improving accuracy across a wide range of conditions, including quiet speech, abrupt audio boundaries and repetitive dialogue patterns that previously caused errors.
The result is a transcription system that is not only more accurate with more challenging media content, but also more adaptable as new networks and content types are onboarded. Benchmarking showed that Google’s Gemini consistently outperformed other large language models for WBD’s use case, while Google Cloud offered the flexibility to run Gemini alongside in-house and open-source models, all of which operate in a closed-loop environment, ensuring WBD’s IP and data are protected and used only by and for WBD.
The work paid off. The tool is being used with select, unscripted content, and captioning that once took hours can now be finished in minutes. “Our core KPI is cost per minute of content processed while maintaining the same high-quality bar,” Dr Devkar says, “and that includes both the automation cost—model inference, compute infrastructure—and the human-in-the-loop cost.”
Humans stay in the loop
WBD tracks average edit time per asset and the number of edits per title as indicators of model quality, linking human and AI performance. “Technology and operations have to evolve together,” Dr Devkar emphasises. “It’s not a simple software rollout; it’s a workflow transformation.” Editorial teams now review AI-generated captions in Spanish and Chinese in near real-time as part of a streamlined workflow—one that complements, rather than replaces, existing vendor partnerships as in-house capabilities expand.

The initial focus on a small number of core languages provided a blueprint for broader rollout, allowing WBD to refine both the technical models and operational workflows before expanding further.
Rapid innovation in AI adds another layer of complexity. New models and updates emerge so quickly that deciding when to build versus wait has become a recurring strategic question. In response, WBD has constructed a benchmarking framework with ground-truth datasets and evaluation processes to test and compare new models. “Each of these challenges reshaped our roadmap,” Dr Devkar says. “It made us think more systemically about data, infrastructure and the human layer that has to evolve alongside the technology.”

It’s about democratising our rich catalogue and enabling as many fans as possible around the world to connect with our content in their language of choice.”
Deepna Devkar
Senior vice president of machine learning and artificial intelligence for global streaming at WBD
For now, Dr Devkar’s attention is on connecting systems that have historically operated in isolation. “We’re already using AI to help audiences discover content in a meaningful way,” she says. “The next step is linking those systems together—so that insights from distribution and engagement inform one another.” Such feedback loops can help predict which stories will resonate in different markets, tailor marketing more effectively and support teams with data-driven insights earlier in the process.
For a company with a century of material and an increasingly international audience, linking those systems matters as much as accelerating them. The infrastructure now in place leaves WBD better equipped to adapt as global viewing habits continue to evolve.
References
- 1.YPulse, “Survey: Younger Viewers Embrace Subtitles”, August 21st 2023.
