new language models

New language models released

Continuous development
At Zoom Media we’re always working on improvement of existing models and developing new ones. For two very different use cases we’ve now deployed two new language models: a Dutch language model specifically designed for call center content and a Filipino model for broadcast content.

Transcribing phone calls
Different use cases ask for different speech recognition models. Zoom Media originally focussed solely on developing speech recognition models for the media industry (hence the name!), meaning the models could be trained with high quality (broadcast) content in order to build 16khz language models. Phone calls are a whole different story however. Due to compression and subsequent loss of important data needed for a high accuracy, phone calls require a custom build 8khz model. Together with partners we have trained our generic Dutch 16khz model to handle these phone calls the right way and the results are impressive. Now call centers and help desks are able to transcribe all their audio content the same way media companies already could, meaning with the same high level of accuracy that Zoom Media stands for.

Filipino for the deaf and hard of hearing
In a previous blog I already mentioned the increasing need for public companies to make their content, apps, programmes and so forth, better accessible for people who are deaf or hard of hearing. Digital accessibility is already regulatory in the Netherlands for example for city councils, provinces, parliament and public broadcasters are taking first steps towards fulfilling their obligation as well. It’s not just in Europe that digital accessibility is an important matter. On the other side of the globe, similar legislation has been adopted by the Filipino government. Broadcasters are obligated, as a start, to automatically subtitle 5-6 hours of content per day in 2019, a number which will gradually increase over the years to come. In light of this legislation Zoom Media has worked together with Microsoft to develop a Filipino language model that works in real-time and can thus meet the requirements set by the government. The model has been deployed, reaching an accuracy of 90% with clear speech, as do all Zoom Media models, enabling broadcasters in the Philippines to accurately subtitle their content in real-time. Read more about digital accessibility in the Philippines here, try out Filipino yourself over here, or drop me an e-mail to discuss how we can make your content better accessible.

Share this post

Share on twitter
Share on linkedin
Share on email