 |
| IGTV gets automatic captions |
At the time of publication, Facebook announced that it would provide automatic IGTV subtitles in up to 16 languages worldwide.
This version was released after the launch of Facebook Live and Workplace Live's automatic translation feature, which was released in March in six languages: English, Spanish, Portuguese, Italian, German and French.
According to Facebook, the improved translation feature is based on an alternate text update that was made a few years ago to help people with low vision.
"As more people start using closed captions and learning artificial intelligence, we hope the quality will continue to improve," she added. It's a small step and we're looking to expand it to other places, languages and countries. "
The social platform mentioned in the blog that it has developed machine learning model training techniques that support automatic speech recognition to directly predict the lettering of words, making it easier to train the model.
Engineers trained the model to adapt to the new words and used public Facebook posts to predict where the new words would be in the video.
Facebook also said it has been able to deliver these models through several infrastructure improvements to serve the additional video traffic generated by the pandemic's load.
According to the platform, the number of live broadcasts on Facebook pages doubled in June 2020 compared to the same period last year.
Facebook launched its first product with automatic translation for video ads in February 2016. In October of the same year, Facebook introduced a free tool to place video tags in English on all Facebook pages.
Although the tool has improved over the years, evidence shows it still has a long way to go.
As Forbes pointed out in a recent article, translation errors can seriously affect the listening experience of people with impaired hearing.
Given a clear understanding of the flaws in the system, Facebook looked for ways to improve translation in the future.
In a study last month, Facebook tried a new method to reduce word error rates (a commonly used indicator of speech recognition performance) by more than 20%.