Mozilla to Represent Speech Synthesis System LPCNet

LPCNet uses DSP for LPC filtering (Linear Predictive Coding) and voice path modeling
21 November 2018   504

Mozilla reported on the new speech synthesis system LPCNet, which effectively translates text to speech while reducing resource demands. This is achieved through a combination of traditional digital signal processing methods (DSP, digital signal processing) with speech synthesis mechanisms based on a recurrent neural network.

The main problem of modern systems for real-time speech synthesis based on neural networks is high computational complexity. It does not allow to use them on smartphones and tablets.

LPCNet uses DSP for LPC filtering (Linear Predictive Coding) and voice path modeling. Then, instead of all the selected samples, the neural network receives only the forecast of each subsequent one. This frees the AI ​​from modeling the vocal tract and leaves it with only an adjustment to the problems in forecasting. Neural networks need only to monitor the accuracy of the forecast, and not to generate each sample in real time.

The technology can be used in other areas where you need to improve the quality of the voice signal. For example, for transmitting speech over low-speed communication channels, eliminating noise, filtering data and restoring fragments of speech lost during transmission.

LPCNet is written in C using a high-level framework for building Keras neural networks. A GTX 1080 Ti video card is desirable for operation. Ready-made models are available for download, but the system can be trained on your own data. LPCNet is distributed under the BSD license.

Mozilla's speech synthesis system is being developed as an alternative to WaveNet by Google. The WaveNet code was open to developers in March 2018.

Nvidia to Open StyleGan Source Code

This machine learning project allows to create of people faces by imitating photographs
11 February 2019   777

NVIDIA has open source code if developments related to the StyleGAN project, which allows generating images of new faces of people by imitating photographs. The system automatically takes into account aspects of the placement of individuals and makes the result indistinguishable from real photos (most of the respondents could not distinguish the original photos from the generated ones). For the synthesis of individuals, a machine learning system based on a generative-competitive neural network (GAN) is used. The code is written in Python using the TensorFlow framework and published under the Creative Commons BY-NC 4.0 license (for non-commercial use only).

Both ready-made trained models and collections of images for self-learning of a neural network are available for download. The basic model was trained on the basis of the Flickr-Faces-HQ (FFHQ) collection, which includes 70,000 high-quality (1024x1024) PNG images of people's faces. At the same time, the system is not tied to persons - as an example, the variants trained on collections of photographs of cars, cats and beds are shown. It requires one or more NVIDIA graphics cards (Tesla V100 GPU recommended), at least 11 GB of RAM, NVIDIA 391.35+ drivers, CUDA 9.0+ tools and the cuDNN 7.3.1 library.

The system allows you to synthesize the image of a new face based on interpolation of features of several faces, combining features characteristic of them, as well as adapting the final image to the required age, gender, hair length, smile character, nose shape, skin color, glasses, face rotation in the photo. The generator considers the image as a collection of styles, automatically separates the characteristic details (freckles, hair, glasses) from common high-level attributes (posture, gender, age changes) and allows you to combine them in an arbitrary form with the definition of the dominant properties through weights.