Jean-Marc Valin has been applying deep learning frameworks to audio over the past few years. So far he’s released RNNoise (a surprisingly good/fast denoising system) and LPCNet (a speech synthesis system along the lines of WaveNet, but fast enough to use realtime on commodity hardware).
Now he’s built a codec out of LPCNet.
“A Real-Time Wideband Neural Vocoder at 1.6 kb/s Using LPCNet” presents a new wideband speech codec built out of the best parts of a brutally speed and space efficient vocoder paired with deep-learning analysis and excitation. It’s alpha-grade research in a lot of ways, but decidedly not vapourware. You can download the source and play with it now, but first, go have a look at the demo page.
