neural network

Vocal Mimicry Using Lyrebird Technology

Lyrebird has created a voice imitation technology that uses deep learning and artificial neural networks to create fascinating and somewhat scary results. It relies on deep learning models developed at the MILA lab of the University of Montréal.

Lyrebird will offer an API to copy the voice of anyone. It will need as little as one minute of audio recording of a speaker to compute a unique key defining her/his voice. This key will then allow to generate anything from its corresponding voice. The API will be robust enough to learn from noisy recordings. Lyrebird will offer a large catalog of different voices and let the user design their own unique voices tailored for their needs.

Users will be able to create entire dialogs with the new or mimicked voice. Inflection, emotion, and content can all be tailored as necessary through a developer API. The demos are fairly impressive but still distinctly robotic. Check out some Trump/Obama examples below.

[soundcloud url=”https://api.soundcloud.com/playlists/318022504″ params=”auto_play=false&hide_related=false&show_comments=true&show_user=true&show_reposts=true&visual=true” width=”100%” height=”300″ iframe=”true” /]

I’m interested to find out how accurate this will be for non-english and non-human verbal communication.

Zoom In. Now… Enhance! (For Real, Kinda)

The Zoom And Enhance trope has long been the ultimate criminal identification solution and a staple for crime drama television. Its use on screen is often lauded as an example of how Hollywood doesn’t understand technology. The Enhance Button trope simply ignores that the blurry focus and big blocky pixels you get when you zoom in close on an image are the only information that the picture actually contains, and attempting to extract more detail from the image alone is essentially impossible.

Enhance Old Station

Enhance Bank Lobby

However, as a proof of concept, Alex J. Champandard’s Neural Enhance coding project uses deep learning to enhance the details of images. As seen from the gifs above, if the neural networks are well trained, the enhancements are quite effective.

Thanks to deep learning and #NeuralEnhance, it’s now possible to train a neural network to zoom into your images at 2x or even 4x. You’ll get even better results by increasing the number of neurons or training with a dataset similar to your low-resolution image. The catch? The neural network is hallucinating details based on its training from example images. It’s not reconstructing your photo exactly as it would have been if it was HD. That’s only possible in Holywood — but using deep learning as “Creative AI” works and it’s just as cool!

Now let’s vector in and enlarge the z-axis.

via prosthetic knowledge

Does A Deep Neural Network Like Your #Selfie?

Best Selfies
Andrej Karpathy trained a Convolutional Neural Network with a dataset of 2 million photographs to determine what makes the perfect selfie. The image above contains the top 100 best selfies (here are the 1,000 best selfies) Andrej concludes that the best selfies have these qualities:

  1. Be female. Women are consistently ranked higher than men. In particular, notice that there is not a single guy in the top 100.
  2. Face should occupy about 1/3 of the image. Notice that the position and pose of the face is quite consistent among the top images. The face always occupies about 1/3 of the image, is slightly tilted, and is positioned in the center and at the top. Which also brings me to:
  3. Cut off your forehead. What’s up with that? It looks like a popular strategy, at least for women.
  4. Show your long hair. Notice the frequent prominence of long strands of hair running down the shoulders.
  5. Oversaturate the face. Notice the frequent occurrence of over-saturated lighting, which often makes the face look much more uniform and faded out. Related to that,
  6. Put a filter on it. Black and White photos seem to do quite well, and most of the top images seem to contain some kind of a filter that fades out the image and decreases the contrast.
  7. Add a border. You will notice a frequent appearance of horizontal/vertical white borders.

Andrej also created a TwitterBot that will judge your selfie. Simply attach your selfie (or a include a link) to a tweet that mentions @deepselfie anywhere in it. The bot will analyze your selfie and give you its opinion (e.g. score 90% means that the Selfie Bot is 90% sure yours would be in top half of selfies. Selfie Bot was not impressed with my selfie.


More here.

Scroll to Top