Princeton and Adobe researchers built a system that lets you edit spoken words by changing text. Cut, copy, or synthesize new phrases in the original voice. A blind study found listeners often couldn't tell the edits from real speech.
Dear fellow scholars, this is two Minute Papers with Karoojol Nay Faher.
Dear fellow scholars, this is two Minute Papers with Close enough. As you have probably noticed, today we are going to talk about Text to Speech or tts. In short, TTS means that we write a piece of text and the computer synthesized voice will read it aloud for us. This is really useful for reading the news or creating audiobooks that don't have any official voiceovers. This work was done by researchers at Princeton University and Adobe and is about text based audio narration editing. This one is going to be crazy good. The Adobe guys like to call this the Photoshop of voiceovers. In a normal situation we have access to a waveform and if we wish to change anything in a voiceover we need to edit it. Editing waveforms by hand is extremely difficult. Traditional techniques often can't even reliably find the boundaries between words and letters, let alone edit them. And with this technique we can cut, copy and even edit this text and the waveforms will automatically be transformed appropriately using the same voice.
Had it struck squarely, it would have killed him. Had it struck squarely, it would have saved him.
We can even use new words that have never been uttered in the original narration.
We leave the eventuality to time and law. We leave the eventuality to time and belief. Believe. Belief. Belief. Belief. Belief. Belief. Belief. Belief. We leave the eventuality to time and belief. We leave the eventuality to time and belief.
It solves an optimization problem where the similarity, smoothness and the pace of the original footage is to be matched as closely as possible. One of the excellent new features is that we can even choose from several different voicings for the new word and insert the one that we deem the most appropriate for expert users. The pitch and duration is also editable. It is always important to have a look at a new technique and make sure that it works well in practice. But in science this is only the first step. There has to be more proof that a new proposed method works well in a variety of cases. In this case, a theoretical proof by means of mathematics is not feasible. Therefore, a user study was carried out where listeners were shown synthesized and real audio samples and had to blindly decide which was which. The algorithm was remarkably successful at deceiving the test subjects. Make sure to have a look at the paper in the description for more details. This technique is traditional in a sense that it doesn't use any sort of neural networks. However, there are great strides being made in that area as well, which I quite excited to show you in future episodes. And due to some of these newer video and audio editing techniques, I expect that within the Internet forums, fake news is going to be an enduring topic. I hope that in parallel with better and better text and video synthesis, there will be an arms race with other methods that are designed to identify these cases. A neural detective, if you will. And now, if you excuse me, I'll give this publicly available TTS one more try and see if I can retire from narrating videos.
Thanks for watching and for your generous support. And I'll see you next time.
Yep, exact same thing. Bet you didn't even notice it.
