Your voice can make you say anything, without you even knowing it. That is how you could summarise Vall-E, Microsoft’s recently launched text-to-speech tool. After hearing a sample of someone’s voice lasting just 3 seconds, this A.I. model can reproduce spoken sentences almost flawlessly, making it difficult to tell that it is a machine speaking to you rather than the person whose voice it has ‘borrowed’. It can even add intonation and emotions such as anger, frustration and ‘sleepiness’ to the spoken text.
Vall-E has been trained using 60,000 hours of English-language speech recordings.
Vall-E has been trained using 60,000 hours of English-language speech recordings, researchers write in a research paper. The Vall-E model was tested by students at Cornell University, who published a website with various demos. This page features several real speech recordings that were used as samples for Vall-E. For each sample, speech recordings generated by Vall-E are then published.
Read this article in full
Become a Werf& member for unlimited access to all articles, research and exclusive insights into the labour market.
No obligations Cancel anytime Secure payment