Researchers at Microsoft have revealed details about their new speech recognition system, promising that its able to transcribe conversational speech as well as a proper human being.
But, as we all know, us meatbags are pretty flawed. That means the system isn't perfect, and has an error rate of around 5.9 per cent. That's roughly the same as the professional transcribers who worked on the same collection of conversations. It would be nice to have 100 per cent accuracy, but I know I'll take 94.4 per cent over 50 per cent any day.
The new system uses neural language models that process conversation more efficiently by grouping similar words together. The idea being that in the near future it will be implemented into the Cortana virtual assistant and other speech-to-text software.
But as big a step as this is, there's always more work to be done. The tech does need some fine-tuning to ensure it works with a wider variety of conversations and voice types. After all, what good is super-accurate voice recognition if it can't understand your regional accent?
Implementing the tech into Cortana will also have its challenges, ensuring that the system can understand words rather than just being able to write them down.