BBC launches machine-translated synthetic voiceovers
Mon 21 Dec 2015
The BBC has announced the launch of a new production tool to provide audio in alternate languages for its news outputs, apparently incorporating existing translation technologies.
Built by BBC news labs, the workflow involves uploading the script of a video news item and the subsequent voice-synthesis of the resulting translation. The service is launching initially in Japanese and Russian. The video, embedded below, shows the process in action, with the narration provided by one of the synthetic voices – and even if the Hawking-style choppiness gives the simulant away immediately, it does appear to provide an above-averagely authentic flow of speech.
Digital Development Director James Montgomery comments in the post “I’m very excited about this trial. The BBC has some of the best original journalism in the world, with correspondents around the globe. Technology like this means we can bring more of our international journalism to more people.”
The initial pilot will run until April of 2016, with the Japanese iteration already available at bbc.jp.
It’s not clear which of the available translation technologies – such as Google Translate, Reverso or Dictionary.com – is incorporated into the tool, but since it’s well known that none of them are perfect, the report outlines that BBC journalists have to do a fair amount of clearing-up after getting the initial translation back.
Accurate language translation is amongst the hardest challenges facing the developers of algorithms and AIs, since example-based approaches do not always account for idiomatic or uncommon usage, so it seems that it will be quite a while before humans are removed from the process. Microsoft Skype has recently launched a live translation feature which can create subtitles from two sides of a conversation, though it has not announced any UN-style plans to actually machine-speak the translated chat. As an illustration that the BBC is going in at the deep end with Japanese, Skype’s attempts to decipher Chinese inadvertently led to the output of profanity that wasn’t there, in tests earlier in the year.