The Real-Time Wideband Neural Vocoder with 1 . 6 kb/s Using LPCNet

This is an update on the LPCNet project , an effective neural speech synthesizer from Mozilla’ s Emerging Technologies group. Within an an earlier demonstration from late a year ago, we showed how LPCNet mixes signal processing and deep learning how to improve the efficiency of neural conversation synthesis.

This time, we all turn LPCNet into a very low-bitrate neural speech codec that’ ersus actually usable on current equipment and even on phones ( as described in this document ). It’ s the 1st time a neural vocoder is able to operate in real-time using just one PROCESSOR core on a phone (as in opposition to a high-end GPU)! The ensuing bitrate — just 1 . 6  kb/s — is about 10  periods less than what wideband codecs usually use. The quality is much better than current very low bitrate vocoders. In fact , it’ s comparable to that of more traditional codecs using higher bitrates.

Screenshot of a demo player that shows the quality of LPCNet-coded speech

This new codec may be used to improve voice quality in nations with poor network connectivity. It is also used as redundancy to improve strength to packet loss for everyone. Within storage applications, it can compress a good hour-long podcast to just 720  kB (so you’ ll still have space left on your floppy disk). Which includes further work, the technology at the rear of LPCNet could help improve existing codecs at very low bitrates.

Learn more about our ongoing work and look for the playable demo in this article .

If you liked The Real-Time Wideband Neural Vocoder with 1 . 6 kb/s Using LPCNet by Jean-Marc Valin Then you'll love Web Design Agency Miami