This post was originally published at LPCNet: DSP-Boosted Neural Speech Synthesis


LPCNet is a new project out of Mozilla’s Emerging Technologies group — an efficient neural speech synthesiser with reduced complexity over some of its predecessors. Neural speech synthesis models like WaveNet have already demonstrated impressive speech synthesis quality, but their computational complexity has made them hard to use in real-time, especially on phones. In a similar fashion to the RNNoise project, our solution with LPCNet is to use a combination of deep learning and digital signal processing (DSP) techniques.

LPCNet samples screenshot

Figure 1: Screenshot of a demo player that demonstrates the quality of LPCNet-synthesized speech.

LPCNet can help improve the quality of text-to-speech (TTS), low bitrate speech coding, time stretching, and more. You can hear the difference for yourself in our LPCNet demo page, where LPCNet and WaveNet speech are generated with the same complexity. The demo also explains the motivations for LPCNet, shows what it can achieve, and explores its possible applications.

You can find an in-depth explanation of the algorithm used in LPCNet in this paper.

The post LPCNet: DSP-Boosted Neural Speech Synthesis appeared first on Mozilla Hacks – the Web developer blog.

Original article:

This post was originally published at LPCNet: DSP-Boosted Neural Speech Synthesis