DeepMind's WaveNet generates raw audio sample-by-sample using dilated convolutions to synthesize speech that mimics specific voices. It outperforms existing text-to-speech methods and can produce music and non-speech sounds, though the original implementation took 90 minutes to create one second of audio.
Questions this video answers
September 12, 2016.
16 to 24,000 samples per second.
Approximately 90 minutes per 1 second of sound.
Dilated convolutions.
A convolutional neural network (CNN), not a recurrent neural network (RNN).
Chapters
Dear fellow scholars, this is two Minute papers with When I opened my inbox today, I was greeted by a huge deluge of messages about Wavenet. Well, first, it's great to see that so many people are excited about these inventions. And second, may all your wishes come true as quickly as this one. So, here we go. This piece of work is about generating audio waveforms for text to speech. And more than text to speech basically means that we have a voice reading whatever we have written down. The difference in this work is, however, that it can synthesize these samples in someone's voice, provided that we have training samples of this person speaking.
The avocado is a pear shaped fruit with leathery skin, smooth edible flesh and a large stone. The avocado is a pear shaped fruit with leathery skin, smooth edible flesh and a large stone. The avocado is a pear shaped fruit with leathery skin, smooth edible flesh and a large stone. The avocado is a pear shaped fruit with leathery skin, smooth edible flesh and a large stone.
It also generates waveforms sample by sample, which is particularly perilous because we typically need to produce these at the rate of 16 to 24,000 samples per second. And and as we listen to the tv, radio and talk to each other several hours a day. The human ear and brain is particularly suited to processing this kind of signal. If the result is off by only the slightest amount, we immediately recognize it. It is not using a recurrent neural network, which is typically suited to learn sequences of things and is widely used for sound synthesis. It is using a convolutional neural network, which is quite surprising because it is not meant to process sequences data that change in time. However, this variant contains an extension that is able to do that. They call this extension dilated convolutions and they open up the possibility of making large skips in the input data so we have a better global view of it. If we were working in computer vision, it would be like increasing the receptive field of the eye so we can see the entire landscape and not only a tree on a photograph. It is also a bit like the temporal coherence problem we have talked about earlier. Taking all this into consideration results in more consistent outputs over larger timescales, so the technique knows what it had done several seconds ago. Also, training a convolutional neural network is a walk in the park compared to a recurrent neural network. Really cool. And the results beat all existing widely used techniques by a large margin. One of these is the concatenative technique, which builds sentences from a huge amount of small speech fragments. These have seen a ton of improvements during the years, but the outputs are still robotic and it is noticeable that we are not listening to a human, but a computer. The DeepMind guys also report that. Notice that non speech sounds such as breathing and mouth movements are also sometimes generated by Wavenet. This reflects the greater flexibility of a raw audio model.
The Blue lagoon is a 1980American romance and adventure film directed by Randall Kleiser the Blue lagoon is a 1980American romance and adventure film directed by Randall Kleiser Aspects of the Sublime in English Poetry and Painting Aspects of the Sublime in English Poetry and Painting.
At the same time, I'd like to note that in the next few episodes it may be that my voice is a bit different, but don't worry about that. It may also happen that I am on a vacation, but new episodes and voice samples pop up on the channel. Please don't worry about that either. Everything is working as intended. They also experimented with music generation and the results are just stunning. Sa. I don't know what to say. These difficult problems, these impenetrable walls crumble one after another as DeepMind takes on them. Insanity. Their blog post and the paper are both really well written. Make sure to check them out. They are both linked in the video description box. I wager that artistic style transfer for sound and instruments is not only coming, but it will be here soon. I imagine that we'll play a guitar and it will sound like a harp and we'll be able to sing something in Lady Gaga's voice and intonation. I've also seen someone pitching the idea of creating audiobooks automatically with such a technique. Wow. I travel a lot and am almost always on the go so I personally would love to have such audiobooks. I I have linked to the mentioned machine learning Reddit thread in the description box. As always, there's lots of great discussion and ideas there. It was also reported that the algorithm currently takes 90 minutes to synthesize 1 second of sound waveforms. You know the drill. One follow up paper down the line it will take only a few minutes. A few more papers down the line it will be real time. Just think about all these advancements, what a time we are living in and I am extremely excited to present them all to you fellow scholars in two minute papers. Make sure to leave your thoughts and ideas in the comments section. We love reading them. Thanks for watching and for your generous support and I'll see you next time. Sam.
