Her name is Lyra. Use deep learning to analyze the human voice. In 40 seconds, process a 90 minute film and save two days of subtitling work
16.10.2018 • 09:38hs • Software for export
Software for export
Netflix, with local DNA: how Argentina's artificial intelligence works in its premieres
/https://assets.iproup.com/assets/jpg/2018/10/326.jpg)
It is not the lonely bird that consoles itself with singing. But it is called Lyra, in allusion to the bird originally from Australia that is famous for its ability to imitate almost any sound. It is the algorithm capable of processing the audio of a movie or series and returning the position and duration in which the subtitles should be inserted. But the most remarkable thing about this Argentine development is that, since last September, it has been used in Netflix's upcoming releases.
"We use deep learning technology, known as deep learning, which analyzes the human voice. The algorithm learns to detect the segments from many movies and series previously subtitled by human beings. Then, already trained, is able to reproduce this task automatically. The more we teach, the better it works, "explains Ezequiel Ansaldi, founder and CEO of Zowl Labs.
Based in Buenos Aires, the company was born in 2015 in order to specialize in artificial intelligence projects and, in particular, audio and video processing, developing tools to accelerate tedious tasks for human beings.
"The goal of reducing delivery times and, therefore, the cost is achieved because artificial intelligence works without supervision. It's like a robot that assembles parts in a factory: it's cheaper, faster and allows man to take on more complex tasks, "continues the entrepreneur.
In this way, the firm already has among its clients companies of the caliber of Acindar, Andreani, Intellivision, King Agro and Axis Communications, together with research and development groups of the University of Buenos Aires, the Institut National de Recherche en Informatique et en Automatique (INRIA) and the University of Rennes, both of France. "We have more than 40 scientific publications and we are part of the NVIDIA Inception program", adds the entrepreneur.
However, getting to Netflix is, without doubt, a qualitative leap in your initiatives. "The first stage of the project required two months and the work of two developers. But the adjustment of the system is a dynamic process, which continually seeks to retrain the algorithm with new movies and series. Everything points to the fact that, in the future, subtitles can be generated completely automatically and with the translation included. But, at least for now, the expert eye is necessary to validate the result ", Ansaldi completes.
This trained sense, in this case, is borne by the members of Oxobox, a cloud platform created in Buenos Aires, in 2008, which provides audiovisual libraries and collaborative workflows on this type of content.
For example, it manages the commercial distribution of 35% of the largest television advertisers in Latin America. And it supplies the system of digital ad sales to DirecTV in twelve countries in the region. In addition, it contributes to the realization of festivals and advertising prizes in Argentina, Mexico, Colombia and Guatemala. With its different business units, in this way, it invoices about $ 80,000,000 annually.
"Our Software as a Service platform for managing audiovisual material already has more than 2,000 functions, which are in charge from the ingestion and transcoding to the management of licenses and contractual manifestos. But from the technical point of view, we intuited that we could improve our quality control paradigms using artificial intelligence, "says Ramiro Fernández, CEO of Oxobox.
"With our OTT Services division, we are a Netflix Preferred Fulfillment Partner, also called Netflix Preferred Vendor, one of the 27 quality control agents in the world. We have an efficiency reported by Netflix of 99.4%, "he emphasizes.
But as Silicon Valley's on-demand multimedia content streaming company justly demands the continuous improvement of its suppliers' workflows, both in terms of quality and speed of delivery, there could be no better opportunity to implement this solution, as the vanguard of the next generation of complements of the company. In fact, Fernández anticipates that they plan to use artificial intelligence in more than thirty key tasks, focusing on everything that until now was impossible without that tool.
"In the case of subtitles for Netflix, we basically define the functional scope and environmental aspects, and Zowl Labs develops the algorithm. We agree to buy the full development, including the algorithm and the source code without compiling. We reserve all rights of ownership and modification, "Fernandez points out.
The results, meanwhile, exceed expectations. "In 40 seconds, Lyra returns a subtitle without legends, but with precise paragraphs for 90 minutes of content. There is still a lot of retraining to do because, although we have 5,000 hours to feed it, with a set of data that includes 'Luis Miguel, the series', 'The Walking Dead' and 'Clan Story', we must do it with small bites of no more than 100 hours Our idea of a more ambitious use of the algorithm still has a long way to go, but for quality control and forecasting it is already available, "he says.
In fact, it illustrates that those 40 seconds that it takes to process a 90-minute film today allow them to save up to two days in eight-day subtitle subprojects. And, most importantly, according to their perspective, which allows to focus the hiring of human resources according to linguistic and non-technical competences. "This improves the quality of the subtitle and the user experience that attends the content, in addition to lowering the cost and time of the project," he concludes.
:quality(85)/https://assets.iproup.com/assets/jpg/2025/05/42966_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/05/42967_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2024/09/39935_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2022/03/26229_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/05/42952_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/02/41871_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/01/41634_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/05/42960_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/02/41910_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2023/02/33559_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2019/06/4500_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2023/04/34424_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2025/02/41849_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2021/05/18860_landscape.jpg)
:quality(85)/https://assets.iproup.com/assets/jpg/2021/11/24072_landscape.jpg)