Scribbling Speech
Supervised by Prof. Carola Zwick
Bachelor Project, July 2018

Tools:
Google Speech API, Google Natural Language API, Quickdraw Dataset, Processing, After Effects

Idea:
Language and images are closely intertwined: We think in pictures and we explain facts as spatial constellations. What if the spoken word could be transformed into dynamic visual worlds in real time? Speech input, machine learning and recurrent neural networks for image generation allow to com- puter generate complex imaginary worlds that follow the narrator and thus create complex animations controlled by linguistic structures.

Concept Sketch:




Create a Virtual World while Speaking

In the model, the objects have their default position value x, y, and z index. A camera has been applied to open our eyes to see this world.

"I'm going to the house." - Going forward view

"I look to the left/right." - Rotating to the left/right view





Initial Object Position

SkyObject: clouds, sun, moon, airplane, stars...
AirObject: butterfly, bird, mosquito...
FloorObject: house, flowers, building, apple, girl, shoes...

Input - Spoken Words




Real time Voice Recognition and Multi language Detection

With Google Speech API and Google Natural Language API, it is able to extract nouns and adpostions like "above","under" and so on from the real time streaming. RecognizeStream, getEntities and analyzeSyntax are used to realize the input part.

For example. "There is a bird flying above the sea", the computer will return "sea", "bird" and "above" in the end.

Besides, not only English can be recognized, but also many other languages like Chinese, French, German and so on.



Realtime Speech Recognition Demo (Please enable audio)

Realtime Speech Recognition Demo in Chinese

Output - Animated Drawings



Quickdraw Dataset

Quick, Draw! is an AI experiment from Google that has delighted millions of people across the world. The game asks users to draw a doodle, then the game’s AI tries to guess what it is.

The richness and the huge amount of data as well as the simplicity of human doodlings, which contains a common understanding of human depicted common objects. It is a visual language of stereotypes of flower”, “mountain”, “cloud”, “river” etc. Over a billion doodles have been drawn by people playing the game and subsequently collected into this anonymized dataset.

different airplanes drawn by different people in Quickdraw

Creating a sense of hand drawn animation

Process from free speech to animation

Demo




Talking in English (Please enable audio)




Talking in Chinese (Please enable audio)




User interacting with Scribbling Speech




Visualisation of the created 3D space